Metrics to evaluate the performance of auto-segmentation for radiation treatment planning: A critical review

Enhancing Precision in Cardiac Segmentation for Magnetic Resonance-Guided Radiation Therapy Through Deep Learning

PURPOSE

Cardiac substructure dose metrics are more strongly linked to late cardiac morbidities than to whole-heart metrics. Magnetic resonance (MR)-guided radiation therapy (MRgRT) enables substructure visualization during daily localization, allowing potential for enhanced cardiac sparing. We extend a publicly available state-of-the-art deep learning framework, "No New" U-Net, to incorporate self-distillation (nnU-Net.wSD) for substructure segmentation for MRgRT.

METHODS AND MATERIALS

Eighteen (institute A) patients who underwent thoracic or abdominal radiation therapy on a 0.35 T MR-guided linear accelerator were retrospectively evaluated. On each image, 1 of 2 radiation oncologists delineated reference contours of 12 cardiac substructures (chambers, great vessels, and coronary arteries) used to train (n = 10), validate (n = 3), and test (n = 5) nnU-Net.wSD by leveraging a teacher-student network and comparing it to standard 3-dimensional U-Net. The impact of using simulation data or including 3 to 4 daily images for augmentation during training was evaluated for nnU-Net.wSD. Geometric metrics (Dice similarity coefficient, mean distance to agreement, and 95% Hausdorff distance), visual inspection, and clinical dose-volume histograms were evaluated. To determine generalizability, institute A's model was tested on an unlabeled data set from institute B (n = 22) and evaluated via consensus scoring and volume comparisons.

RESULTS

nnU-Net.wSD yielded a Dice similarity coefficient (reported mean ± SD) of 0.65 ± 0.25 across the 12 substructures (chambers, 0.85 ± 0.05; great vessels, 0.67 ± 0.19; and coronary arteries, 0.33 ± 0.16; mean distance to agreement, <3 mm; mean 95% Hausdorff distance, <9 mm) while outperforming the 3-dimensional U-Net (0.583 ± 0.28; P <.01). Leveraging fractionated data for augmentation improved over a single MR simulation time point (0.579 ± 0.29; P <.01). Predicted contours yielded dose-volume histograms that closely matched those of the clinical treatment plans where mean and maximum (ie, dose to 0.03 cc) doses deviated by 0.32 ± 0.5 Gy and 1.42 ± 2.6 Gy, respectively. There were no statistically significant differences between institute A and B volumes (P >.05) for 11 of 12 substructures, with larger volumes requiring minor changes and coronary arteries exhibiting more variability.

CONCLUSIONS

This work is a critical step toward rapid and reliable cardiac substructure segmentation to improve cardiac sparing in low-field MRgRT.

Validation of different automated segmentation models for target volume contouring in postoperative radiotherapy for breast cancer and regional nodal irradiation

Introduction

Target volume delineation is routinely performed in postoperative radiotherapy (RT) for breast cancer patients, but it is a time-consuming process. The aim of the present study was to validate the quality, clinical usability and institutional-specific implementation of different auto-segmentation tools into clinical routine.

Methods

Three different commercially available, artificial intelligence-, ESTRO-guideline-based segmentation models (M1-3) were applied to fifty consecutive reference patients who received postoperative local RT including regional nodal irradiation for breast cancer for the delineation of clinical target volumes: the residual breast, implant or chestwall, axilla levels 1 and 2, the infra- and supraclavicular regions, the interpectoral and internal mammary nodes. Objective evaluation metrics of the created structures were conducted with the Dice similarity index (DICE) and the Hausdorff distance, and a manual evaluation of usability.

Results

The resulting geometries of the segmentation models were compared to the reference volumes for each patient and required no or only minor corrections in 72 % (M1), 64 % (M2) and 78 % (M3) of the cases. The median DICE and Hausdorff values for the resulting planning target volumes were 0.87-0.88 and 2.96-3.55, respectively. Clinical usability was significantly correlated with the DICE index, with calculated cut-off values used to define no or minor adjustments of 0.82-0.86. Right or left sided target and breathing method (deep inspiration breath hold vs. free breathing) did not impact the quality of the resulting structures.

Conclusion

Artificial intelligence-based auto-segmentation programs showed high-quality accuracy and provided standardization and efficient support for guideline-based target volume contouring as a precondition for fully automated workflows in radiotherapy treatment planning.

Impact of non-contrast-enhanced imaging input sequences on the generation of virtual contrast-enhanced breast MRI scans using neural network

OBJECTIVE

To investigate how different combinations of T1-weighted (T1w), T2-weighted (T2w), and diffusion-weighted imaging (DWI) impact the performance of virtual contrast-enhanced (vCE) breast MRI.

MATERIALS AND METHODS

The IRB-approved, retrospective study included 1064 multiparametric breast MRI scans (age: 52 ± 12 years) obtained from 2017 to 2020 (single site, two 3-T MRI). Eleven independent neural networks were trained to derive vCE images from varying input combinations of T1w, T2w, and multi-b-value DWI sequences (b-value = 50-1500 s/mm2). Three readers evaluated the vCE images with regard to qualitative scores of diagnostic image quality, image sharpness, satisfaction with contrast/signal-to-noise ratio, and lesion/non-mass enhancement conspicuity. Quantitative metrics (SSIM, PSNR, NRMSE, and median symmetrical accuracy) were analyzed and statistically compared between the input combinations for the full breast volume and both enhancing and non-enhancing target findings.

RESULTS

The independent test set consisted of 187 cases. The quantitative metrics significantly improved in target findings when multi-b-value DWI sequences were included during vCE training (p < 0.05). Non-significant effects (p > 0.05) were observed for the quantitative metrics on the full breast volume when comparing input combinations including T1w. Using T1w and DWI acquisitions during vCE training is necessary to achieve high satisfaction with contrast/SNR and good conspicuity of the enhancing findings. The input combination of T1w, T2w, and DWI sequences with three b-values showed the best qualitative performance.

CONCLUSION

vCE breast MRI performance is significantly influenced by input sequences. Quantitative metrics and visual quality of vCE images significantly benefit when multi b-value DWI is added to morphologic T1w-/T2w sequences as input for model training.

KEY POINTS

Question How do different MRI sequences impact the performance of virtual contrast-enhanced (vCE) breast MRI? Findings The input combination of T1-weighted, T2-weighted, and diffusion-weighted imaging sequences with three b-values showed the best qualitative performance. Clinical relevance While in the future neural networks providing virtual contrast-enhanced images might further improve accessibility to breast MRI, the significant influence of input data needs to be considered during translational research.

Achieving accurate prostate auto-segmentation on CT in the absence of MR imaging

BACKGROUND

Magnetic resonance imaging (MRI) is considered the gold standard for prostate segmentation. Computed tomography (CT)-based segmentation is prone to observer bias, potentially overestimating the prostate volume by ∼ 30 % compared to MRI. However, MRI accessibility is challenging for patients with contraindications or in rural areas globally with limited clinical resources.

PURPOSE

This study investigates the possibility of achieving MRI-level prostate auto-segmentation accuracy using CT-only input via a deep learning (DL) model trained with CT-MRI registered segmentation.

METHODS AND MATERIALS

A cohort of 111 definitive prostate radiotherapy patients with both CT and MRI images was retrospectively grouped into training (n = 37) and validation (n = 20) (where reference contours were derived from CT-MRI registration), and testing (n = 54) sets. Two commercial DL models were benchmarked against the reference contours in the training and validation sets. A custom DL model was incrementally retrained using the training dataset, quantitatively evaluated on the validation dataset, and qualitatively assessed by two different physician groups on the validation and testing datasets. A contour quality assurance (QA) model, established from the proposed model on the validation dataset, was applied to the test group to identify potential errors, confirmed by human visual inspection.

RESULTS

Two commercial models exhibited large deviations in the prostate apex with CT-only input (median: 0.77/0.78 for Dice similarity coefficient (DSC), and 0.80 cm/0.83 cm for 95 % directed Hausdorff Distance (HD95), respectively). The proposed model demonstrated superior geometric similarity compared to commercial models, particularly in the apex region, with improvements of 0.05/0.17 cm and 0.06/0.25 cm in median DSC/HD95, respectively. Physician evaluation on MRI-CT registration data rated 69 %-78 % of the proposed model's contours as clinically acceptable without modifications. Additionally, 73 % of cases flagged by the contour quality assurance (QA) model were confirmed via visual inspection.

CONCLUSIONS

The proposed incremental learning strategy based on CT-MRI registration information enhances prostate segmentation accuracy when MRI availability is limited clinically.

A comparative study on automatic treatment planning for online adaptive proton therapy of esophageal cancer: which combination of deformable registration and deep learning planning tools performs the best?

Objective.To demonstrate the feasibility of integrating fully-automated online adaptive proton therapy strategies (OAPT) within a commercially available treatment planning system and underscore what limits their clinical implementation. These strategies leverage existing deformable image registration (DIR) algorithms and state-of-the-art deep learning (DL) networks for organ segmentation and proton dose prediction.Approach.Four OAPT strategies featuring automatic segmentation and robust optimization were evaluated on a cohort of 17 patients, each undergoing a repeat CT scan. (1) DEF-INIT combines deformably registered contours with template-based optimization. (2) DL-INIT, (3) DL-DEF, and (4) DL-DL employ a nnU-Net DL network for organ segmentation and a controlling ROIs-guided DIR algorithm for internal clinical target volume (iCTV) segmentation. DL-INIT uses this segmentation alongside template-based optimization, DL-DEF integrates it with a dose-mimicking (DM) step using a reference deformed dose, and DL-DL merges it with DM on a reference DL-predicted dose. All strategies were evaluated on manual contours and contours used for optimization and compared with manually adapted plans. Key dose volume metrics like iCTV D98% are reported.Main results.iCTV D98% was comparable in manually adapted plans and for all strategies in nominal cases but dropped to 20 Gy in worst-case scenarios for a few patients per strategy, highlighting the need to correct segmentation errors in the target volume. Evaluations on optimization contours showed minimal relative error, with some outliers, particularly in template-based strategies (DEF-INIT and DL-INIT). DL-DEF achieves a good trade-off between speed and dosimetric quality, showing a passing rate (iCTV D98% > 94%) of 90% when evaluated against 2, 4 and 5 mm setup error and of 88% when evaluated against 7 mm setup error. While template-based methods are more rigid, DL-DEF and DL-DL have potential for further enhancements with proper DM algorithm tuning.Significance.Among investigated strategies, DL-DEF and DL-DL demonstrated promising within 10 min OAPT implementation results and significant potential for improvements.

Evaluation of the accuracy of automated segmentation based on deep learning for prostate cancer patients

PURPOSE

This study evaluated the accuracy of a commercial deep learning (DL)-based algorithm for segmenting the prostate, seminal vesicles (SV), and organs at risk (OAR) in patients with prostate cancer.

METHODS

Ten patients with prostate cancer were selected to compare automated and manual segmentation. The prostate, SV, and OAR, including the bladder, rectum, left and right femoral heads, and penile bulb, were delineated and reviewed according to our institutional protocols by radiation oncologists. The CT and MR images were fused to the prostate, and the prostate and penile bulb were manually delineated on the CT and MR images. The remaining organs were delineated on the CT images without the MR images. MVision AI Contour+ was used to perform DL-based automated segmentation. The dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95%) were evaluated for comparison with manual delineations.

RESULTS

The mean DSC values for the prostate, SV, bladder, rectum, both femoral heads, and penile bulb were 0.86, 0.80, 0.96, 0.92, 0.97, and 0.64, respectively. The HD95% for all the organs was less than 3 mm.

CONCLUSIONS

The commercial DL-based auto segmentation solution provided high-quality contours in patients with prostate cancer.

Development and comprehensive evaluation of a national DBCG consensus-based auto-segmentation model for lymph node levels in breast cancer radiotherapy

BACKGROUND AND PURPOSE

This study aimed at training and validating a multi-institutional deep learning (DL) auto segmentation model for nodal clinical target volume (CTVn) in high-risk breast cancer (BC) patients with both training and validation dataset created with multi-institutional participation, with the overall aim of national clinical implementation in Denmark.

MATERIALS AND METHODS

A gold standard (GS) dataset and a high-quality training dataset were created by 21 BC delineation experts from all radiotherapy centres in Denmark. The delineations were created according to ESTRO consensus delineation guidelines. Four models were trained: One per laterality and extension of CTVn internal mammary nodes. The DL models were tested quantitatively in their own test-set and in relation to interobserver variation (IOV) in the GS dataset with geometrical metrics, such as the Dice Similarity Coefficient (DSC). A blinded qualitative evaluation was conducted with a national board, presented to both DL and manual delineations.

RESULTS

A median DSC > 0.7 was found for all, except the CTVn interpectoral node in one of the models. In the qualitative evaluation 'no corrections needed' were acquired for 297 (36 %) in the DL structures and 286 (34 %) for manual delineations. A higher rate of 'major corrections' and 'easier to start from scratch' was found in the manual delineations. The models performed within the IOV of an expert group, with two exceptions.

CONCLUSION

DL models were developed on a national consensus cohort and performed on par with the IOV between BC experts and had a comparable or higher clinical acceptance than expert manual delineations.

ASF-LKUNet: Adjacent-scale fusion U-Net with large kernel for multi-organ segmentation

In the multi-organ segmentation task of medical images, there are some challenging issues such as the complex background, blurred boundaries between organs, and the larger scale difference in volume. Due to the local receptive fields of conventional convolution operations, it is difficult to obtain desirable results by directly using them for multi-organ segmentation. While Transformer-based models have global information, there is a significant dependency on hardware because of the high computational demands. Meanwhile, the depthwise convolution with large kernel can capture global information and have less computational requirements. Therefore, to leverage the large receptive field and reduce model complexity, we propose a novel CNN-based approach, namely adjacent-scale fusion U-Net with large kernel (ASF-LKUNet) for multi-organ segmentation. We utilize a u-shaped encoder-decoder as the base architecture of ASF-LKUNet. In the encoder path, we design the large kernel residual block, which combines the large and small kernels and can simultaneously capture the global and local features. Furthermore, for the first time, we propose an adjacent-scale fusion and large kernel GRN channel attention that incorporates the low-level details with the high-level semantics by the adjacent-scale feature and then adaptively focuses on the more global and meaningful channel information. Extensive experiments and interpretability analysis are made on the Synapse multi-organ dataset (Synapse) and the ACDC cardiac multi-structure dataset (ACDC). Our proposed ASF-LKUNet achieves 88.41% and 89.45% DSC scores on the Synapse and ACDC datasets, respectively, with 17.96M parameters and 29.14 GFLOPs. These results show that our method achieves superior performance with favorable lower complexity against ten competing approaches.ASF-LKUNet is superior to various competing methods and has less model complexity. Code and the trained models have been released on GitHub.

Geometric and dosimetric evaluation for breast and regional nodal auto-segmentation structures

The accuracy of artificial intelligence (AI) generated contours for intact-breast and post-mastectomy radiotherapy plans was evaluated. Geometric and dosimetric comparisons were performed between auto-contours (ACs) and manual-contours (MCs) produced by physicians for target structures. Breast and regional nodal structures were manually delineated on 66 breast cancer patients. ACs were retrospectively generated. The characteristics of the breast/post-mastectomy chestwall (CW) and regional nodal structures (axillary [AxN], supraclavicular [SC], internal mammary [IM]) were geometrically evaluated by Dice similarity coefficient (DSC), mean surface distance, and Hausdorff Distance. The structures were also evaluated dosimetrically by superimposing the MC clinically delivered plans onto the ACs to assess the impact of utilizing ACs with target dose (Vx%) evaluation. Positive geometric correlations between volume and DSC for intact-breast, AxN, and CW were observed. Little or anti correlations between volume and DSC for IM and SC were shown. For intact-breast plans, insignificant dosimetric differences between ACs and MCs were observed for AxNV95% (p = 0.17) and SCV95% (p = 0.16), while IMNV90% ACs and MCs were significantly different. The average V95% for intact-breast MCs (98.4%) and ACs (97.1%) were comparable but statistically different (p = 0.02). For post-mastectomy plans, AxNV95% (p = 0.35) and SCV95% (p = 0.08) were consistent between ACs and MCs, while IMNV90% was significantly different. Additionally, 94.1% of AC-breasts met ΔV95% variation <5% when DSC > 0.7. However, only 62.5% AC-CWs achieved the same metrics, despite AC-CWV95% (p = 0.43) being statistically insignificant. The AC intact-breast structure was dosimetrically similar to MCs. The AC AxN and SC may require manual adjustments. Careful review should be performed for AC post-mastectomy CW and IMN before treatment planning. The findings of this study may guide the clinical decision-making process for the utilization of AI-driven ACs for intact-breast and post-mastectomy plans. Before clinical implementation of this auto-segmentation software, an in-depth assessment of agreement with each local facilities MCs is needed.

Autodelineation methods in a simulated fully automated proton therapy workflow for esophageal cancer

Background and purpose

Proton Online Adaptive RadioTherapy (ProtOnART) harnesses the dosimetric advantage of protons and immediately acts upon anatomical changes. Here, we simulate the clinical application of delineation and planning within a ProtOnART-workflow for esophageal cancer. We aim to identify the most appropriate technique for autodelineation and evaluate full automation by replanning on autodelineated contours.

Materials and methods

We evaluated 15 patients who started treatment between 11-2022 and 01-2024, undergoing baseline and three repeat computed tomography (CT) scans in treatment position. Quantitative and qualitative evaluations compared different autodelineation methods. For Organs-at-risk (OAR) deep learning segmentation (DLS), rigid and deformable propagation from baseline to repeat CT-scans were considered. For the clinical target volume (CTV), rigid and three deformable propagation methods (default, heart as controlling structure and with focus region) were evaluated. Adaptive treatment plans with 7 mm (ATP7mm) and 3 mm (ATP3mm) setup robustness were generated using best-performing autodelineated contours. Clinical acceptance of ATPs was evaluated using goals encompassing ground-truth CTV-coverage and OAR-dose.

Results

Deformation was preferred for autodelineation of heart, lungs and spinal cord. DLS was preferred for all other OARs. For CTV, deformation with focus region was the preferred method although the difference with other deformation methods was small. Nominal ATPs passed evaluation goals for 87 % of ATP7mm and 67 % of ATP3mm. This dropped to respectively 2 % and 29 % after robust evaluation. Insufficient CTV-coverage was the main reason for ATP-rejection.

Conclusion

Autodelineation aids a ProtOnART-workflow for esophageal cancer. Currently available tools regularly require manual annotations to generate clinically acceptable ATPs.

Efficient application of deep learning-based elective lymph node regions delineation for pelvic malignancies

BACKGROUND

While there are established international consensuses on the delineation of pelvic lymph node regions (LNRs), significant inter- and intra-observer variabilities persist. Contouring these clinical target volumes for irradiation in pelvic malignancies is both time-consuming and labor-intensive.

PURPOSE

The purpose of this study was to develop a deep learning model of pelvic LNRs delineation for patients with pelvic cancers.

METHODS

Planning computed tomography (CT) studies of 160 patients with pelvic primary malignancies (including rectal, prostate, and cervical cancer) were retrospectively collected and divided into training set (n = 120) and testing set (n = 40). Six pelvic LNRs, including abdominal presacral, pelvic presacral, internal iliac nodes, external iliac nodes, obturator nodes, and inguinal nodes were delineated by two radiation oncologists as ground truth (Gt) contours. The cascaded multi-heads U-net (CMU-net) was constructed based on the Gt contours from training cohort, which was subsequently verified in the testing cohort. The automatic delineation of six LNRs (Auto) was evaluated using dice similarity coefficient (DSC), average surface distance (ASD), 95th percentile Hausdorff distance (HD95), and a 7-point scale score.

RESULTS

In the testing set, the DSC of six pelvic LNRs by CMU-net model varied from 0.851 to 0.942, ASD from 0.381 to 1.037 mm, and HD95 from 2.025 to 3.697 mm. No significant differences were founded in these three parameters between postoperative and preoperative cases. 95.9% and 96.2% of auto delineations by CMU-net model got a score of 1-3 by two expert radiation oncologists, respectively, meaning only minor edits needed.

CONCLUSIONS

The CMU-net was successfully developed for automated delineation of pelvic LNRs for pelvic malignancies radiotherapy with improved contouring efficiency and highly consistent, which might justify its implementation in radiotherapy work flow.

Impact of motion management strategies on abdominal organ at risk delineation for magnetic resonance-guided radiotherapy

Background and purpose

The impact of respiratory motion management strategies for abdominal radiotherapy, such as abdominal compression (AC) and breath hold (BH), on abdominal organ at risk (OAR) delineation on magnetic resonance imaging (MRI) is unknown. This feasibility study compared the inter- and intra- observer delineation variation on MRI acquired with AC, BH for three critical abdominal OAR.

Materials and methods

T2-weighted (W) 3D MRI in free-breathing (FB) and with AC, and T1W 3D mDixon exhale BH were acquired. Four observers blinded to motion management strategy used, delineated stomach, liver, and duodenum on all MRI. One case per strategy was repeated over 6 weeks later to quantify intra-observer variation. Simultaneous truth and performance level estimation (STAPLE) contours for each OAR were generated, median and IQR mean distance to agreement (mDTA) and maximum Hausdorff distance (HD) between observer and STAPLE contours were calculated. Observers scored organ visibility on each MRI using a four-point Likert scale.

Results

A total of 27 scans including repeats were delineated. Pooled mDTA for all OARs was 1.3 mm (0.5 mm) with AC, 1.4 mm (1.0 mm) with BH, and 1.3 mm (0.5 mm) in FB. Intra-observer mDTA was highest for all organs in FB with 10.8 mm for duodenum, 1.8 mm for liver, and 2.7 mm for stomach. The pooled mean perceptual quality score value was highest for AC across organs.

Conclusions

No motion management strategy demonstrated superior similarity across OAR, emphasizing the need for personalised approaches based on individual clinical and patient factors.

Comprehensive Clinical Usability-Oriented Contour Quality Evaluation for Deep Learning Auto-segmentation: Combining Multiple Quantitative Metrics Through Machine Learning

PURPOSE

The current commonly used metrics for evaluating the quality of auto-segmented contours have limitations and do not always reflect the clinical usefulness of the contours. This work aims to develop a novel contour quality classification (CQC) method by combining multiple quantitative metrics for clinical usability-oriented contour quality evaluation for deep learning-based auto-segmentation (DLAS).

METHODS AND MATERIALS

The CQC was designed to categorize contours on slices as acceptable, minor edit, or major edit based on the expected editing effort/time with supervised ensemble tree classification models using 7 quantitative metrics. Organ-specific models were trained for 5 abdominal organs (pancreas, duodenum, stomach, small, and large bowels) using 50 magnetic resonance imaging (MRI) data sets. Twenty additional MRI and 9 computed tomography (CT) data sets were employed for testing. Interobserver variation (IOV) was assessed among 6 observers and consensus labels were established through majority vote for evaluation. The CQC was also compared with a threshold-based baseline approach.

RESULTS

For the 5 organs, the average area under the curve was 0.982 ± 0.01 and 0.979 ± 0.01, the mean accuracy was 95.8% ± 1.7% and 94.3% ± 2.1%, and the mean risk rate was 0.8% ± 0.4% and 0.7% ± 0.5% for MRI and CT testing data set, respectively. The CQC results closely matched the IOV results (mean accuracy of 94.2% ± 0.8% and 94.8% ± 1.7%) and were significantly higher than those obtained using the threshold-based method (mean accuracy of 80.0% ± 4.7%, 83.8% ± 5.2%, and 77.3% ± 6.6% using 1, 2, and 3 metrics).

CONCLUSIONS

The CQC models demonstrated high performance in classifying the quality of contour slices. This method can address the limitations of existing metrics and offers an intuitive and comprehensive solution for clinically oriented evaluation and comparison of DLAS systems.

Custom-Trained Deep Learning-Based Auto-Segmentation for Male Pelvic Iterative CBCT on C-Arm Linear Accelerators

PURPOSE

The purpose of this investigation was to evaluate the clinical applicability of a commercial artificial intelligence-driven deep learning auto-segmentation (DLAS) tool on enhanced iterative cone beam computed tomography (iCBCT) acquisitions for intact prostate and prostate bed treatments.

METHODS AND MATERIALS

DLAS models were trained using 116 iCBCT data sets with manually delineated organs at risk (bladder, femoral heads, and rectum) and target volumes (intact prostate and prostate bed) adhering to institution-specific contouring guidelines. An additional 25 intact prostate and prostate bed iCBCT data sets were used for model testing. Segmentation accuracy relative to a reference structure set was quantified using various geometric comparison metrics and qualitatively evaluated by trained physicists and physicians. These results were compared with those obtained for an additional DLAS-based model trained on planning computed tomography (pCT) data sets and for a deformable image registration (DIR)-based automatic contour propagation method.

RESULTS

In most instances, statistically significant differences in the Dice similarity coefficient (DSC), 95% directed Hausdorff distance, and mean surface distance metrics were observed between the models, as the iCBCT-trained DLAS model outperformed the pCT-trained DLAS model and DIR-based method for all organs at risk and the intact prostate target volume. Mean DSC values for the proposed method were ≥0.90 for these volumes of interest. The iCBCT-trained DLAS model demonstrated a relatively suboptimal performance for the prostate bed segmentation, as the mean DSC value was <0.75 for this target contour. Overall, 90% of bladder, 93% of femoral head, 67% of rectum, and 92% of intact prostate contours generated by the proposed method were deemed clinically acceptable based on qualitative scoring, and approximately 63% of prostate bed contours required moderate or major manual editing to adhere to institutional contouring guidelines.

CONCLUSIONS

The proposed method presents the potential for improved segmentation accuracy and efficiency compared with the DIR-based automatic contour propagation method as commonly applied in CBCT-based dose evaluation and calculation studies.

Deep learning for autosegmentation for radiotherapy treatment planning: State-of-the-art and novel perspectives

The rapid development of artificial intelligence (AI) has gained importance, with many tools already entering our daily lives. The medical field of radiation oncology is also subject to this development, with AI entering all steps of the patient journey. In this review article, we summarize contemporary AI techniques and explore the clinical applications of AI-based automated segmentation models in radiotherapy planning, focusing on delineation of organs at risk (OARs), the gross tumor volume (GTV), and the clinical target volume (CTV). Emphasizing the need for precise and individualized plans, we review various commercial and freeware segmentation tools and also state-of-the-art approaches. Through our own findings and based on the literature, we demonstrate improved efficiency and consistency as well as time savings in different clinical scenarios. Despite challenges in clinical implementation such as domain shifts, the potential benefits for personalized treatment planning are substantial. The integration of mathematical tumor growth models and AI-based tumor detection further enhances the possibilities for refining target volumes. As advancements continue, the prospect of one-stop-shop segmentation and radiotherapy planning represents an exciting frontier in radiotherapy, potentially enabling fast treatment with enhanced precision and individualization.

Radiomics as a measure superior to common similarity metrics for tumor segmentation performance evaluation

PURPOSE

To propose radiomics features as a superior measure for evaluating the segmentation ability of physicians and auto-segmentation tools and to compare its performance with the most commonly used metrics: Dice similarity coefficient (DSC), surface Dice similarity coefficient (sDSC), and Hausdorff distance (HD).

MATERIALS/METHODS

The data of 10 lung cancer patients' CT images with nine tumor segmentations per tumor were downloaded from the RIDER (Reference Database to Evaluate Response) database. Radiomics features of 90 segmented tumors were extracted using the PyRadiomics program. The intraclass correlation coefficient (ICC) of radiomics features were used to evaluate the segmentation similarity and compare their performance with DSC, sDSC, and HD. We calculated one ICC per radiomics feature and per tumor for nine segmentations and 36 ICCs per radiomics feature for 36 pairs of nine segmentations. Meanwhile, there were 360 DSC, sDSC, and HD values calculated for 36 pairs for 10 tumors.

RESULTS

The ICC of radiomics features exhibited greater sensitivity to segmentation changes than DSC and sDSC. The ICCs of the wavelet-LLL first order Maximum, wavelet-LLL glcm MCC, wavelet-LLL glcm Cluster Shade features ranged from 0.130 to 0.997, 0.033 to 0.978, and 0.160 to 0.998, respectively. On the other hand, all DSC and sDSC were larger than 0.778 and 0.700, respectively, while HD varied from 0 to 1.9 mm. The results indicated that the radiomics features could capture subtle variations in tumor segmentation characteristics, which could not be easily detected by DSC and sDSC.

CONCLUSIONS

This study demonstrates the superiority of radiomics features with ICC as a measure for evaluating a physician's tumor segmentation ability and the performance of auto-segmentation tools. Radiomics features offer a more sensitive and comprehensive evaluation, providing valuable insights into tumor characteristics. Therefore, the new metrics can be used to evaluate new auto-segmentation methods and enhance trainees' segmentation skills in medical training and education.

Analyzing the Relationship between Dose and Geometric Agreement Metrics for Auto-Contouring in Head and Neck Normal Tissues

This study aimed to determine the relationship between geometric and dosimetric agreement metrics in head and neck (H&N) cancer radiotherapy plans. A total 287 plans were retrospectively analyzed, comparing auto-contoured and clinically used contours using a Dice similarity coefficient (DSC), surface DSC (sDSC), and Hausdorff distance (HD). Organs-at-risk (OARs) with ≥200 cGy dose differences from the clinical contour in terms of Dmax (D0.01cc) and Dmean were further examined against proximity to the planning target volume (PTV). A secondary set of 91 plans from multiple institutions validated these findings. For 4995 contour pairs across 19 OARs, 90% had a DSC, sDSC, and HD of at least 0.75, 0.86, and less than 7.65 mm, respectively. Dosimetrically, the absolute difference between the two contour sets was <200 cGy for 95% of OARs in terms of Dmax and 96% in terms of Dmean. In total, 97% of OARs exhibiting significant dose differences between the clinically edited contour and auto-contour were within 2.5 cm PTV regardless of geometric agreement. There was an approximately linear trend between geometric agreement and identifying at least 200 cGy dose differences, with higher geometric agreement corresponding to a lower fraction of cases being identified. Analysis of the secondary dataset validated these findings. Geometric indices are approximate indicators of contour quality and identify contours exhibiting significant dosimetric discordance. For a small subset of OARs within 2.5 cm of the PTV, geometric agreement metrics can be misleading in terms of contour quality.

Artificial intelligence in radiotherapy: Current applications and future trends

Radiation therapy has dramatically changed with the advent of computed tomography and intensity modulation. This added complexity to the workflow but allowed for more precise and reproducible treatment. As a result, these advances required the accurate delineation of many more volumes, raising questions about how to delineate them, in a uniform manner across centers. Then, as computing power improved, reverse planning became possible and three-dimensional dose distributions could be generated. Artificial intelligence offers the opportunity to make such workflow more efficient while increasing practice homogeneity. Many artificial intelligence-based tools are being implemented in routine practice to increase efficiency, reduce workload and improve homogeneity of treatments. Data retrieved from this workflow could be combined with clinical data and omic data to develop predictive tools to support clinical decision-making process. Such predictive tools are at the stage of proof-of-concept and need to be explainatory, prospectively validated, and based on large and multicenter cohorts. Nevertheless, they could bridge the gap to personalized radiation oncology, by personalizing oncologic strategies, dose prescriptions to tumor volumes and dose constraints to organs at risk.

The current landscape of machine learning-based radiomics in arteriovenous malformations: a systematic review and radiomics quality score assessment

Background

Arteriovenous malformations (AVMs) are rare vascular anomalies involving a disorganization of arteries and veins with no intervening capillaries. In the past 10 years, radiomics and machine learning (ML) models became increasingly popular for analyzing diagnostic medical images. The goal of this review was to provide a comprehensive summary of current radiomic models being employed for the diagnostic, therapeutic, prognostic, and predictive outcomes in AVM management.

Methods

A systematic literature review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, in which the PubMed and Embase databases were searched using the following terms: (cerebral OR brain OR intracranial OR central nervous system OR spine OR spinal) AND (AVM OR arteriovenous malformation OR arteriovenous malformations) AND (radiomics OR radiogenomics OR machine learning OR artificial intelligence OR deep learning OR computer-aided detection OR computer-aided prediction OR computer-aided treatment decision). A radiomics quality score (RQS) was calculated for all included studies.

Results

Thirteen studies were included, which were all retrospective in nature. Three studies (23%) dealt with AVM diagnosis and grading, 1 study (8%) gauged treatment response, 8 (62%) predicted outcomes, and the last one (8%) addressed prognosis. No radiomics model had undergone external validation. The mean RQS was 15.92 (range: 10-18).

Conclusion

We demonstrated that radiomics is currently being studied in different facets of AVM management. While not ready for clinical use, radiomics is a rapidly emerging field expected to play a significant future role in medical imaging. More prospective studies are warranted to determine the role of radiomics in the diagnosis, prediction of comorbidities, and treatment selection in AVM management.

Application of simultaneous uncertainty quantification and segmentation for oropharyngeal cancer use-case with Bayesian deep learning

BACKGROUND

Radiotherapy is a core treatment modality for oropharyngeal cancer (OPC), where the primary gross tumor volume (GTVp) is manually segmented with high interobserver variability. This calls for reliable and trustworthy automated tools in clinician workflow. Therefore, accurate uncertainty quantification and its downstream utilization is critical.

METHODS

Here we propose uncertainty-aware deep learning for OPC GTVp segmentation, and illustrate the utility of uncertainty in multiple applications. We examine two Bayesian deep learning (BDL) models and eight uncertainty measures, and utilize a large multi-institute dataset of 292 PET/CT scans to systematically analyze our approach.

RESULTS

We show that our uncertainty-based approach accurately predicts the quality of the deep learning segmentation in 86.6% of cases, identifies low performance cases for semi-automated correction, and visualizes regions of the scans where the segmentations likely fail.

CONCLUSIONS

Our BDL-based analysis provides a first-step towards more widespread implementation of uncertainty quantification in OPC GTVp segmentation.

Gross failure rates and failure modes for a commercial AI-based auto-segmentation algorithm in head and neck cancer patients

PURPOSE

Artificial intelligence (AI) based commercial software can be used to automatically delineate organs at risk (OAR), with potential for efficiency savings in the radiotherapy treatment planning pathway, and reduction of inter- and intra-observer variability. There has been little research investigating gross failure rates and failure modes of such systems.

METHOD

50 head and neck (H&N) patient data sets with "gold standard" contours were compared to AI-generated contours to produce expected mean and standard deviation values for the Dice Similarity Coefficient (DSC), for four common H&N OARs (brainstem, mandible, left and right parotid). An AI-based commercial system was applied to 500 H&N patients. AI-generated contours were compared to manual contours, outlined by an expert human, and a gross failure was set at three standard deviations below the expected mean DSC. Failures were inspected to assess reason for failure of the AI-based system with failures relating to suboptimal manual contouring censored. True failures were classified into 4 sub-types (setup position, anatomy, image artefacts and unknown).

RESULTS

There were 24 true failures of the AI-based commercial software, a gross failure rate of 1.2%. Fifteen failures were due to patient anatomy, four were due to dental image artefacts, three were due to patient position and two were unknown. True failure rates by OAR were 0.4% (brainstem), 2.2% (mandible), 1.4% (left parotid) and 0.8% (right parotid).

CONCLUSION

True failures of the AI-based system were predominantly associated with a non-standard element within the CT scan. It is likely that these non-standard elements were the reason for the gross failure, and suggests that patient datasets used to train the AI model did not contain sufficient heterogeneity of data. Regardless of the reasons for failure, the true failure rate for the AI-based system in the H&N region for the OARs investigated was low (∼1%).

Associations Between Radiation Oncologist Demographic Factors and Segmentation Similarity Benchmarks: Insights From a Crowd-Sourced Challenge Using Bayesian Estimation

PURPOSE

The quality of radiotherapy auto-segmentation training data, primarily derived from clinician observers, is of utmost importance. However, the factors influencing the quality of clinician-derived segmentations are poorly understood; our study aims to quantify these factors.

METHODS

Organ at risk (OAR) and tumor-related segmentations provided by radiation oncologists from the Contouring Collaborative for Consensus in Radiation Oncology data set were used. Segmentations were derived from five disease sites: breast, sarcoma, head and neck (H&N), gynecologic (GYN), and GI. Segmentation quality was determined on a structure-by-structure basis by comparing the observer segmentations with an expert-derived consensus, which served as a reference standard benchmark. The Dice similarity coefficient (DSC) was primarily used as a metric for the comparisons. DSC was stratified into binary groups on the basis of structure-specific expert-derived interobserver variability (IOV) cutoffs. Generalized linear mixed-effects models using Bayesian estimation were used to investigate the association between demographic variables and the binarized DSC for each disease site. Variables with a highest density interval excluding zero were considered to substantially affect the outcome measure.

RESULTS

Five hundred seventy-four, 110, 452, 112, and 48 segmentations were used for the breast, sarcoma, H&N, GYN, and GI cases, respectively. The median percentage of segmentations that crossed the expert DSC IOV cutoff when stratified by structure type was 55% and 31% for OARs and tumors, respectively. Regression analysis revealed that the structure being tumor-related had a substantial negative impact on binarized DSC for the breast, sarcoma, H&N, and GI cases. There were no recurring relationships between segmentation quality and demographic variables across the cases, with most variables demonstrating large standard deviations.

CONCLUSION

Our study highlights substantial uncertainty surrounding conventionally presumed factors influencing segmentation quality relative to benchmarks.

From quantitative metrics to clinical success: assessing the utility of deep learning for tumor segmentation in breast surgery

PURPOSE

Preventing positive margins is essential for ensuring favorable patient outcomes following breast-conserving surgery (BCS). Deep learning has the potential to enable this by automatically contouring the tumor and guiding resection in real time. However, evaluation of such models with respect to pathology outcomes is necessary for their successful translation into clinical practice.

METHODS

Sixteen deep learning models based on established architectures in the literature are trained on 7318 ultrasound images from 33 patients. Models are ranked by an expert based on their contours generated from images in our test set. Generated contours from each model are also analyzed using recorded cautery trajectories of five navigated BCS cases to predict margin status. Predicted margins are compared with pathology reports.

RESULTS

The best-performing model using both quantitative evaluation and our visual ranking framework achieved a mean Dice score of 0.959. Quantitative metrics are positively associated with expert visual rankings. However, the predictive value of generated contours was limited with a sensitivity of 0.750 and a specificity of 0.433 when tested against pathology reports.

CONCLUSION

We present a clinical evaluation of deep learning models trained for intraoperative tumor segmentation in breast-conserving surgery. We demonstrate that automatic contouring is limited in predicting pathology margins despite achieving high performance on quantitative metrics.

Deep Learning Method for Rapid Simultaneous Multistructure Temporal Bone Segmentation

OBJECTIVE

To develop and validate a deep learning algorithm for the automated segmentation of key temporal bone structures from clinical computed tomography (CT) data sets.

STUDY DESIGN

Cross-sectional study.

SETTING

A total of 325 CT scans from a clinical database.

METHOD

A state-of-the-art deep learning (DL) algorithm (SwinUNETR) was used to train a prediction model for rapid segmentation of 9 key temporal bone structures in a data set of 325 clinical CTs. The data set was manually annotated by a specialist to serve as the ground truth. The data set was randomly split into training (n = 260) and testing (n = 65) sets. The model's performance was objectively assessed through external validation on the test set using metrics including Dice, Balanced accuracy, Hausdorff distances, and processing time.

RESULTS

The model achieved an average Dice coefficient of 0.87 for all structures, an average balanced accuracy of 0.94, an average Hausdorff distance of 0.79 mm, and an average processing time of 9.1 seconds per CT.

CONCLUSION

The present DL model for the automated simultaneous segmentation of multiple structures within the temporal bone from CTs achieved high accuracy according to currently commonly employed objective analysis. The results demonstrate the potential of the method to improve preoperative evaluation and intraoperative guidance in otologic surgery.

Enabling Navigation and Augmented Reality in the Sitting Position in Posterior Fossa Surgery Using Intraoperative Ultrasound

Despite its broad use in cranial and spinal surgery, navigation support and microscope-based augmented reality (AR) have not yet found their way into posterior fossa surgery in the sitting position. While this position offers surgical benefits, navigation accuracy and thereof the use of navigation itself seems limited. Intraoperative ultrasound (iUS) can be applied at any time during surgery, delivering real-time images that can be used for accuracy verification and navigation updates. Within this study, its applicability in the sitting position was assessed. Data from 15 patients with lesions within the posterior fossa who underwent magnetic resonance imaging (MRI)-based navigation-supported surgery in the sitting position were retrospectively analyzed using the standard reference array and new rigid image-based MRI-iUS co-registration. The navigation accuracy was evaluated based on the spatial overlap of the outlined lesions and the distance between the corresponding landmarks in both data sets, respectively. Image-based co-registration significantly improved (p < 0.001) the spatial overlap of the outlined lesion (0.42 ± 0.30 vs. 0.65 ± 0.23) and significantly reduced (p < 0.001) the distance between the corresponding landmarks (8.69 ± 6.23 mm vs. 3.19 ± 2.73 mm), allowing for the sufficient use of navigation and AR support. Navigated iUS can therefore serve as an easy-to-use tool to enable navigation support for posterior fossa surgery in the sitting position.

Cardiac substructure delineation in radiation therapy - A state-of-the-art review

Delineation of cardiac substructures is crucial for a better understanding of radiation-related cardiotoxicities and to facilitate accurate and precise cardiac dose calculation for developing and applying risk models. This review examines recent advancements in cardiac substructure delineation in the radiation therapy (RT) context, aiming to provide a comprehensive overview of the current level of knowledge, challenges and future directions in this evolving field. Imaging used for RT planning presents challenges in reliably visualising cardiac anatomy. Although cardiac atlases and contouring guidelines aid in standardisation and reduction of variability, significant uncertainties remain in defining cardiac anatomy. Coupled with the inherent complexity of the heart, this necessitates auto-contouring for consistent large-scale data analysis and improved efficiency in prospective applications. Auto-contouring models, developed primarily for breast and lung cancer RT, have demonstrated performance comparable to manual contouring, marking a significant milestone in the evolution of cardiac delineation practices. Nevertheless, several key concerns require further investigation. There is an unmet need for expanding cardiac auto-contouring models to encompass a broader range of cancer sites. A shift in focus is needed from ensuring accuracy to enhancing the robustness and accessibility of auto-contouring models. Addressing these challenges is paramount for the integration of cardiac substructure delineation and associated risk models into routine clinical practice, thereby improving the safety of RT for future cancer patients.

NRG Oncology Assessment of Artificial Intelligence Deep Learning-Based Auto-segmentation for Radiation Therapy: Current Developments, Clinical Considerations, and Future Directions

Deep learning neural networks (DLNN) in Artificial intelligence (AI) have been extensively explored for automatic segmentation in radiotherapy (RT). In contrast to traditional model-based methods, data-driven AI-based models for auto-segmentation have shown high accuracy in early studies in research settings and controlled environment (single institution). Vendor-provided commercial AI models are made available as part of the integrated treatment planning system (TPS) or as a stand-alone tool that provides streamlined workflow interacting with the main TPS. These commercial tools have drawn clinics' attention thanks to their significant benefit in reducing the workload from manual contouring and shortening the duration of treatment planning. However, challenges occur when applying these commercial AI-based segmentation models to diverse clinical scenarios, particularly in uncontrolled environments. Contouring nomenclature and guideline standardization has been the main task undertaken by the NRG Oncology. AI auto-segmentation holds the potential clinical trial participants to reduce interobserver variations, nomenclature non-compliance, and contouring guideline deviations. Meanwhile, trial reviewers could use AI tools to verify contour accuracy and compliance of those submitted datasets. In recognizing the growing clinical utilization and potential of these commercial AI auto-segmentation tools, NRG Oncology has formed a working group to evaluate the clinical utilization and potential of commercial AI auto-segmentation tools. The group will assess in-house and commercially available AI models, evaluation metrics, clinical challenges, and limitations, as well as future developments in addressing these challenges. General recommendations are made in terms of the implementation of these commercial AI models, as well as precautions in recognizing the challenges and limitations.

Dosimetric impact of contour editing on CT and MRI deep-learning autosegmentation for brain OARs

PURPOSE

To establish the clinical applicability of deep-learning organ-at-risk autocontouring models (DL-AC) for brain radiotherapy. The dosimetric impact of contour editing, prior to model training, on performance was evaluated for both CT and MRI-based models. The correlation between geometric and dosimetric measures was also investigated to establish whether dosimetric assessment is required for clinical validation.

METHOD

CT and MRI-based deep learning autosegmentation models were trained using edited and unedited clinical contours. Autosegmentations were dosimetrically compared to gold standard contours for a test cohort. D1%, D5%, D50%, and maximum dose were used as clinically relevant dosimetric measures. The statistical significance of dosimetric differences between the gold standard and autocontours was established using paired Student's t-tests. Clinically significant cases were identified via dosimetric headroom to the OAR tolerance. Pearson's Correlations were used to investigate the relationship between geometric measures and absolute percentage dose changes for each autosegmentation model.

RESULTS

Except for the right orbit, when delineated using MRI models, the dosimetric statistical analysis revealed no superior model in terms of the dosimetric accuracy between the CT DL-AC models or between the MRI DL-AC for any investigated brain OARs. The number of patients where the clinical significance threshold was exceeded was higher for the optic chiasm D1% than other OARs, for all autosegmentation models. A weak correlation was consistently observed between the outcomes of dosimetric and geometric evaluations.

CONCLUSIONS

Editing contours before training the DL-AC model had no significant impact on dosimetry. The geometric test metrics were inadequate to estimate the impact of contour inaccuracies on dose. Accordingly, dosimetric analysis is needed to evaluate the clinical applicability of DL-AC models in the brain.

Technical note: The migration distance - a unidirectional distance metric for region-of-interest comparisons

BACKGROUND

The radiotherapy process relies on several metrics in determining a notion of "distance" from one three-dimensional region-of-interest (ROI) to another. The majority are symmetric (or commutative) and do not contain information pertaining to directionality. Growth versus regression, for example, is not inherently distinguished by these metrics.

PURPOSE

The purpose of this work was to formalize a unidirectional distance metric, motivated by radiotherapy margin concepts, which we term the migration distance. Informally, the migration distance from ROI X $X$ to Y $Y$ is the minimum isotropic expansion of X $X$ such that Y $Y$ is completely encompassed by the expansion. If Y $Y$ is contained within X $X$ , the migration distance is negative with magnitude equal to the maximum isotropic contraction of X $X$ such that Y $Y$ remains contained within contraction. The metric is demonstrated by quantifying glioblastoma interfraction target changes.

METHODS

An explicit mathematical formulation of the migration distance is presented and contrasted with the related Hausdorff distance. The results are demonstrated for the gross tumor volume (GTV) dynamics of a glioblastoma cohort consisting of 111 patients that underwent standard chemoradiotherapy with offline MR imaging at planning, fraction 10, fraction 20, and 1-month post radiotherapy.

RESULTS

The mean ± SD of the GTV migration distance relative to planning was 5.9 ± 3.9 mm at fraction 10, 6.2 ± 4.4 mm at fraction 20, and 7.9 ± 7.1 mm at 1-month post radiotherapy. The maximum GTV migration distance across all patients at the same timepoints was 20.4, 20.7, and 45.5 mm, respectively.

CONCLUSIONS

We have proposed and demonstrated a unidirectional distance metric. The migration distance may have applications in the quantification of anatomical changes, planning target volume designs, and dosimetric radiotherapy plan assessment.

Large-scale dose evaluation of deep learning organ contours in head-and-neck radiotherapy by leveraging existing plans

Background and purpose

Retrospective dose evaluation for organ-at-risk auto-contours has previously used small cohorts due to additional manual effort required for treatment planning on auto-contours. We aimed to do this at large scale, by a) proposing and assessing an automated plan optimization workflow that used existing clinical plan parameters and b) using it for head-and-neck auto-contour dose evaluation.

Materials and methods

Our automated workflow emulated our clinic's treatment planning protocol and reused existing clinical plan optimization parameters. This workflow recreated the original clinical plan (P OG ) with manual contours (P MC ) and evaluated the dose effect (P OG - P MC ) on 70 photon and 30 proton plans of head-and-neck patients. As a use-case, the same workflow (and parameters) created a plan using auto-contours (P AC ) of eight head-and-neck organs-at-risk from a commercial tool and evaluated their dose effect (P MC - P AC ).

Results

For plan recreation (P OG - P MC ), our workflow had a median impact of 1.0% and 1.5% across dose metrics of auto-contours, for photon and proton respectively. Computer time of automated planning was 25% (photon) and 42% (proton) of manual planning time. For auto-contour evaluation (P MC - P AC ), we noticed an impact of 2.0% and 2.6% for photon and proton radiotherapy. All evaluations had a median Δ NTCP (Normal Tissue Complication Probability) less than 0.3%.

Conclusions

The plan replication capability of our automated program provides a blueprint for other clinics to perform auto-contour dose evaluation with large patient cohorts. Finally, despite geometric differences, auto-contours had a minimal median dose impact, hence inspiring confidence in their utility and facilitating their clinical adoption.

A Prospective Study Measuring Resident and Faculty Contour Concordance: A Potential Tool for Quantitative Assessment of Residents' Performance in Contouring and Target Delineation in Radiation Oncology Residency

PURPOSE/OBJECTIVE(S)

Accurate target delineation (ie, contouring) is essential for radiation treatment planning and radiotherapy efficacy. As a result, improving the quality of target delineation is an important goal in the education of radiation oncology residents. The purpose of this study was to track the concordance of radiation oncology residents' contours with those of faculty physicians over the course of 1 year to assess for patterns.

MATERIALS/METHODS

Residents in postgraduate year (PGY) levels 2 to 4 were asked to contour target volumes that were then compared to the finalized, faculty physician-approved contours. Concordance between resident and faculty physician contours was determined by calculating the Jaccard concordance index (JCI), ranging from 0, meaning no agreement, to 1, meaning complete agreement. Multivariate mixed-effect models were used to assess the association of JCI to the fixed effect of PGY level and its interactions with cancer type and other baseline characteristics. Post hoc means of JCI were compared between PGY levels after accounting for multiple comparisons using Tukey's method.

RESULTS

In total, 958 structures from 314 patients collected during the 2020-2021 academic year were studied. The mean JCI was 0.77, 0.75, and 0.61 for the PGY-4, PGY-3, and PGY-2 levels, respectively. The JCI score for PGY-2 was found to be lower than those for PGY-3 and PGY-4, respectively (all P < .001). No statistically significant difference of JCI score was found between the PGY-3 and PGY-4 levels. The average JCI score was lowest (0.51) for primary head and/or neck cancers, and it was highest (0.80) for gynecologic cancers.

CONCLUSIONS

Tracking and comparing the concordance of resident contours with faculty physician contours is an intriguing method of assessing resident performance in contouring and target delineation and could potentially serve as a quantitative metric, which is lacking currently, in radiation oncology resident evaluation. However, additional study is necessary before this technique can be incorporated into residency assessments.

Artificial Intelligence-Based Autosegmentation: Advantages in Delineation, Absorbed Dose-Distribution, and Logistics

Purpose

The study's purpose was to compare the performance of artificial intelligence (AI) in auto-contouring compared with a human practitioner in terms of precision, differences in dose distribution, and time consumption.

Methods and Materials

Datasets of previously irradiated patients in 3 different segments (head and neck, breast, and prostate cancer) were retrospectively collected. An experienced radiation oncologist (MD) performed organs-at-risk (OARs) and standard clinical target volume delineations as baseline structures for comparison. AI-based autocontours were generated in 2 additional CT copies; therefore, 3 groups were assessed: MD alone, AI alone, and AI plus MD corrections (AI+C). Differences in Dice similarity coefficient (DSC) and person-hour burden were assessed. Furthermore, changes in clinically relevant dose-volume parameters were evaluated and compared.

Results

Seventy-five previously treated cases were collected (25 per segment) for the analysis. Compared with MD contours, the mean DSC scores were higher than 0.7 for 74% and 80% of AI and AI+C, respectively. After corrections, 17.1% structures presented DSC score deviations higher than 0.1 and 10.4% dose-volume parameters significantly changed in AI-contoured structures. The time consumption assessment yielded mean person-hour reductions of 68%, 51%, and 71% for breast, prostate, and head and neck cancer, respectively.

Conclusions

In great extent, AI yielded clinically acceptable OARs and certain clinical target volumes in the explored anatomic segments. Sparse correction and assessment requirements place AI+C as a standard workflow. Minimal clinically relevant differences in OAR exposure were identified. A substantial amount of person-hours could be repurposed with this technology.

Essentially unedited deep-learning-based OARs are suitable for rigorous oropharyngeal and laryngeal cancer treatment planning

Quality of organ at risk (OAR) autosegmentation is often judged by concordance metrics against the human-generated gold standard. However, the ultimate goal is the ability to use unedited autosegmented OARs in treatment planning, while maintaining the plan quality. We tested this approach with head and neck (HN) OARs generated by a prototype deep-learning (DL) model on patients previously treated for oropharyngeal and laryngeal cancer. Forty patients were selected, with all structures delineated by an experienced physician. For each patient, a set of 13 OARs were generated by the DL model. Each patient was re-planned based on original targets and unedited DL-produced OARs. The new dose distributions were then applied back to the manually delineated structures. The target coverage was evaluated with inhomogeneity index (II) and the relative volume of regret. For the OARs, Dice similarity coefficient (DSC) of areas under the DVH curves, individual DVH objectives, and composite continuous plan quality metric (PQM) were compared. The nearly identical primary target coverage for the original and re-generated plans was achieved, with the same II and relative volume of regret values. The average DSC of the areas under the corresponding pairs of DVH curves was 0.97 ± 0.06. The number of critical DVH points which met the clinical objectives with the dose optimized on autosegmented structures but failed when evaluated on the manual ones was 5 of 896 (0.6%). The average OAR PQM score with the re-planned dose distributions was essentially the same when evaluated either on the autosegmented or manual OARs. Thus, rigorous HN treatment planning is possible with OARs segmented by a prototype DL algorithm with minimal, if any, manual editing.

Assessing the Feasibility of Using Artificial Intelligence-Segmented Dominant Intraprostatic Lesion for Focal Intraprostatic Boost With External Beam Radiation Therapy

PURPOSE

The delineation of dominant intraprostatic gross tumor volumes (GTVs) on multiparametric magnetic resonance imaging (mpMRI) can be subject to interobserver variability. We evaluated whether deep learning artificial intelligence (AI)-segmented GTVs can provide a similar degree of intraprostatic boosting with external beam radiation therapy (EBRT) as radiation oncologist (RO)-delineated GTVs.

METHODS AND MATERIALS

We identified 124 patients who underwent mpMRI followed by EBRT between 2010 and 2013. A reference GTV was delineated by an RO and approved by a board-certified radiologist. We trained an AI algorithm for GTV delineation on 89 patients, and tested the algorithm on 35 patients, each with at least 1 PI-RADS (Prostate Imaging Reporting and Data System) 4 or 5 lesion (46 total lesions). We then asked 5 additional ROs to independently delineate GTVs on the test set. We compared lesion detectability and geometric accuracy of the GTVs from AI and 5 ROs against the reference GTV. Then, we generated EBRT plans (77 Gy prostate) that boosted each observer-specific GTV to 95 Gy. We compared reference GTV dose (D98%) across observers using a mixed-effects model.

RESULTS

On a lesion level, AI GTV exhibited a sensitivity of 82.6% and positive predictive value of 86.4%. Respective ranges among the 5 RO GTVs were 84.8% to 95.7% and 95.1% to 100.0%. Among 30 GTVs mutually identified by all observers, no significant differences in Dice coefficient were detected between AI and any of the 5 ROs. Across all patients, only 2 of 5 ROs had a reference GTV D98% that significantly differed from that of AI by 2.56 Gy (P = .02) and 3.20 Gy (P = .003). The presence of false-negative (-5.97 Gy; P < .001) but not false-positive (P = .24) lesions was associated with reference GTV D98%.

CONCLUSIONS

AI-segmented GTVs demonstrate potential for intraprostatic boosting, although the degree of boosting may be adversely affected by false-negative lesions. Prospective review of AI-segmented GTVs remains essential.

Investigation of autosegmentation techniques on T2-weighted MRI for off-line dose reconstruction in MR-linac workflow for head and neck cancers

BACKGROUND

In order to accurately accumulate delivered dose for head and neck cancer patients treated with the Adapt to Position workflow on the 1.5T magnetic resonance imaging (MRI)-linear accelerator (MR-linac), the low-resolution T2-weighted MRIs used for daily setup must be segmented to enable reconstruction of the delivered dose at each fraction.

PURPOSE

In this pilot study, we evaluate various autosegmentation methods for head and neck organs at risk (OARs) on on-board setup MRIs from the MR-linac for off-line reconstruction of delivered dose.

METHODS

Seven OARs (parotid glands, submandibular glands, mandible, spinal cord, and brainstem) were contoured on 43 images by seven observers each. Ground truth contours were generated using a simultaneous truth and performance level estimation (STAPLE) algorithm. Twenty total autosegmentation methods were evaluated in ADMIRE: 1-9) atlas-based autosegmentation using a population atlas library (PAL) of 5/10/15 patients with STAPLE, patch fusion (PF), random forest (RF) for label fusion; 10-19) autosegmentation using images from a patient's 1-4 prior fractions (individualized patient prior [IPP]) using STAPLE/PF/RF; 20) deep learning (DL) (3D ResUNet trained on 43 ground truth structure sets plus 45 contoured by one observer). Execution time was measured for each method. Autosegmented structures were compared to ground truth structures using the Dice similarity coefficient, mean surface distance (MSD), Hausdorff distance (HD), and Jaccard index (JI). For each metric and OAR, performance was compared to the inter-observer variability using Dunn's test with control. Methods were compared pairwise using the Steel-Dwass test for each metric pooled across all OARs. Further dosimetric analysis was performed on three high-performing autosegmentation methods (DL, IPP with RF and 4 fractions [IPP_RF_4], IPP with 1 fraction [IPP_1]), and one low-performing (PAL with STAPLE and 5 atlases [PAL_ST_5]). For five patients, delivered doses from clinical plans were recalculated on setup images with ground truth and autosegmented structure sets. Differences in maximum and mean dose to each structure between the ground truth and autosegmented structures were calculated and correlated with geometric metrics.

RESULTS

DL and IPP methods performed best overall, all significantly outperforming inter-observer variability and with no significant difference between methods in pairwise comparison. PAL methods performed worst overall; most were not significantly different from the inter-observer variability or from each other. DL was the fastest method (33 s per case) and PAL methods the slowest (3.7-13.8 min per case). Execution time increased with a number of prior fractions/atlases for IPP and PAL. For DL, IPP_1, and IPP_RF_4, the majority (95%) of dose differences were within ± 250 cGy from ground truth, but outlier differences up to 785 cGy occurred. Dose differences were much higher for PAL_ST_5, with outlier differences up to 1920 cGy. Dose differences showed weak but significant correlations with all geometric metrics (R2 between 0.030 and 0.314).

CONCLUSIONS

The autosegmentation methods offering the best combination of performance and execution time are DL and IPP_1. Dose reconstruction on on-board T2-weighted MRIs is feasible with autosegmented structures with minimal dosimetric variation from ground truth, but contours should be visually inspected prior to dose reconstruction in an end-to-end dose accumulation workflow.

Clinical evaluation of deep learning and atlas-based auto-segmentation for organs at risk delineation

Manual delineation of organs at risk and clinical target volumes is essential in radiotherapy planning. Atlas-based auto-segmentation (ABAS) algorithms have become available and been shown to provide accurate contouring for various anatomical sites. Recently, deep learning auto-segmentation (DL-AS) algorithms have emerged as the state-of-the-art in medical image segmentation. This study aimed to evaluate the effect of auto-segmentation on the clinical workflow for contouring different anatomical sites of cancer, such as head and neck (H&N), breast, abdominal region, and prostate. Patients with H&N, breast, abdominal, and prostate cancer (n = 30 each) were enrolled in the study. Twenty-seven different organs at four sites were evaluated. RayStation was used to apply the ABAS. Siemens AI-Rad Companion Organs RT was used to apply the DL-AS. Evaluations were performed with similarity indices using geometric methods, time-evaluation, and qualitative scoring visual evaluations by radiation oncologists. The DL-AS algorithm was more accurate than ABAS algorithm on geometric indices for half of the structures. The qualitative scoring results of the two algorithms were significantly different, and DL-AS was more accurate on many contours. DL-AS had 41%, 29%, 86%, and 15% shorter edit times in the HnN, breast, abdomen, and prostate groups, respectively, than ABAS. There were no correlations between the geometric indices and visual assessments. The time required to edit the contours was considerably shorter for DL-AS than for ABAS. Auto-segmentation with deep learning could be the first step for clinical workflow optimization in radiotherapy.

Assessment of residual geometrical errors of clinical target volumes and their impact on dose accumulation for head and neck radiotherapy

PURPOSE

To assess the residual geometrical errors (dr) and their impact on the clinical target volumes (CTV) dose coverage for head and neck cancer (HNC) proton therapy patients.

METHODS

We analysed 28 HNC patients treated with 70 Gy (RBE) and 54.25 Gy (RBE) to the therapeutic CTV70 and prophylactic CTV54.25, respectively. Daily cone beam CTs were converted to high quality synthetic CTs (sCTs). The CTVs from the nominal CT were propagated to the corresponding sCTs using a hybrid deformable image registration (propagated CTVs) in RayStation 11B. For 11 patients, all propagated CTVs were reviewed by our HNC radiation oncologist (physician corrected CTVs). The residual geometrical error dr was quantified as a function of the daily CTVs volume overlap with the nominal plan CTV. The errors dr(propagated CTVs) and dr(physician corrected CTVs) and the difference in dice similarity coefficients (ΔDSC) were determined. Using clinical plans, dose coverage and the tumor control probability (TCP) for the nominal, accumulated and voxel-wise minimum scenarios were determined.

RESULTS

The difference in the residual geometrical error dr (propagated CTVs - physician corrected CTVs) and mean DSC (|ΔDSC|mean) were minor: Δdr(CTV70) = 0.16 mm, Δdr(CTV54.25) = 0.26 mm, |ΔDSC|mean < 0.9%. For all 28 patients, dr(CTV70) = 1.91 mm and dr(CTV54.25) = 1.90 mm. However, CTV54.25 above and below the cricoid cartilage differed substantially (1.00 mm c.f. 3.93 mm). The CTV54.25 coverage below the cricoid was then almost always lower, although the TCP of the accumulated dose was higher than the TCP of the voxel-wise minimum dose.

CONCLUSIONS

Setup uncertainty setting of 2 mm is possible. The feasibility of using propagated CTVs for error determination is demonstrated.

Development and implementation of optimized endogenous contrast sequences for delineation in adaptive radiotherapy on a 1.5T MR-linear-accelerator: a prospective R-IDEAL stage 0-2a quantitative/qualitative evaluation of in vivo site-specific quality-assurance using a 3D T2 fat-suppressed platform for head and neck cancer

Purpose

To improve segmentation accuracy in head and neck cancer (HNC) radiotherapy treatment planning for the 1.5T hybrid magnetic resonance imaging/linear accelerator (MR-Linac), three-dimensional (3D), T2-weighted, fat-suppressed magnetic resonance imaging sequences were developed and optimized.

Approach

After initial testing, spectral attenuated inversion recovery (SPAIR) was chosen as the fat suppression technique. Five candidate SPAIR sequences and a nonsuppressed, T2-weighted sequence were acquired for five HNC patients using a 1.5T MR-Linac. MR physicists identified persistent artifacts in two of the SPAIR sequences, so the remaining three SPAIR sequences were further analyzed. The gross primary tumor volume, metastatic lymph nodes, parotid glands, and pterygoid muscles were delineated using five segmentors. A robust image quality analysis platform was developed to objectively score the SPAIR sequences on the basis of qualitative and quantitative metrics.

Results

Sequences were analyzed for the signal-to-noise ratio and the contrast-to-noise ratio and compared with fat and muscle, conspicuity, pairwise distance metrics, and segmentor assessments. In this analysis, the nonsuppressed sequence was inferior to each of the SPAIR sequences for the primary tumor, lymph nodes, and parotid glands, but it was superior for the pterygoid muscles. The SPAIR sequence that received the highest combined score among the analysis categories was recommended to Unity MR-Linac users for HNC radiotherapy treatment planning.

Conclusions

Our study led to two developments: an optimized, 3D, T2-weighted, fat-suppressed sequence that can be disseminated to Unity MR-Linac users and a robust image quality analysis pathway that can be used to objectively score SPAIR sequences and can be customized and generalized to any image quality optimization protocol. Improved segmentation accuracy with the proposed SPAIR sequence will potentially lead to improved treatment outcomes and reduced toxicity for patients by maximizing the target coverage and minimizing the radiation exposure of organs at risk.

Evaluation of different algorithms for automatic segmentation of head-and-neck lymph nodes on CT images

PURPOSE

To investigate the performance of 4 atlas-based (multi-ABAS) and 2 deep learning (DL) solutions for head-and-neck (HN) elective nodes (CTVn) automatic segmentation (AS) on CT images.

MATERIAL AND METHODS

Bilateral CTVn levels of 69 HN cancer patients were delineated on contrast-enhanced planning CT. Ten and 49 patients were used for atlas library and for training a mono-centric DL model, respectively. The remaining 20 patients were used for testing. Additionally, three commercial multi-ABAS methods and one commercial multi-centric DL solution were investigated. Quantitative evaluation was assessed using volumetric Dice Similarity Coefficient (DSC) and 95-percentile Hausdorff distance (HD95%). Blind evaluation was performed for 3 solutions by 4 physicians. One recorded the time needed for manual corrections. A dosimetric study was finally conducted using automated planning.

RESULTS

Overall DL solutions had better DSC and HD95% results than multi-ABAS methods. No statistically significant difference was found between the 2 DL solutions. However, the contours provided by multi-centric DL solution were preferred by all physicians and were also faster to correct (1.1 min vs 4.17 min, on average). Manual corrections for multi-ABAS contours took on average 6.52 min Overall, decreased contour accuracy was observed from CTVn2 to CTVn3 and to CTVn4. Using the AS contours in treatment planning resulted in underdosage of the elective target volume.

CONCLUSION

Among all methods, the multi-centric DL method showed the highest delineation accuracy and was better rated by experts. Manual corrections remain necessary to avoid elective target underdosage. Finally, AS contours help reducing the workload of manual delineation task.

Interobserver agreement on definition of the target volume in stereotactic radiotherapy for pancreatic adenocarcinoma using different imaging modalities

PURPOSE

The aim of this study was to evaluate interobserver agreement (IOA) on target volume definition for pancreatic cancer (PACA) within the Radiosurgery and Stereotactic Radiotherapy Working Group of the German Society of Radiation Oncology (DEGRO) and to identify the influence of imaging modalities on the definition of the target volumes.

METHODS

Two cases of locally advanced PACA and one local recurrence were selected from a large SBRT database. Delineation was based on either a planning 4D CT with or without (w/wo) IV contrast, w/wo PET/CT, and w/wo diagnostic MRI. Novel compared to other studies, a combination of four metrics was used to integrate several aspects of target volume segmentation: the Dice coefficient (DSC), the Hausdorff distance (HD), the probabilistic distance (PBD), and the volumetric similarity (VS).

RESULTS

For all three GTVs, the median DSC was 0.75 (range 0.17-0.95), the median HD 15 (range 3.22-67.11) mm, the median PBD 0.33 (range 0.06-4.86), and the median VS was 0.88 (range 0.31-1). For ITVs and PTVs the results were similar. When comparing the imaging modalities for delineation, the best agreement for the GTV was achieved using PET/CT, and for the ITV and PTV using 4D PET/CT, in treatment position with abdominal compression.

CONCLUSION

Overall, there was good GTV agreement (DSC). Combined metrics appeared to allow a more valid detection of interobserver variation. For SBRT, either 4D PET/CT or 3D PET/CT in treatment position with abdominal compression leads to better agreement and should be considered as a very useful imaging modality for the definition of treatment volumes in pancreatic SBRT. Contouring does not appear to be the weakest link in the treatment planning chain of SBRT for PACA.

Extensive clinical testing of Deep Learning Segmentation models for thorax and breast cancer radiotherapy planning

BACKGROUND

The performance of deep learning segmentation (DLS) models for automatic organ extraction from CT images in the thorax and breast regions was investigated. Furthermore, the readiness and feasibility of integrating DLS into clinical practice were addressed by measuring the potential time savings and dosimetric impact.

MATERIAL AND METHODS

Thirty patients referred to radiotherapy for breast cancer were prospectively included. A total of 23 clinically relevant left- and right-sided organs were contoured manually on CT images according to ESTRO guidelines. Next, auto-segmentation was executed, and the geometric agreement between the auto-segmented and manually contoured organs was qualitatively assessed applying a scale in the range [0-not acceptable, 3-no corrections]. A quantitative validation was carried out by calculating Dice coefficients (DSC) and the 95% percentile of Hausdorff distances (HD95). The dosimetric impact of optimizing the treatment plans on the uncorrected DLS contours, was investigated from a dose coverage analysis using DVH values of the manually delineated contours as references.

RESULTS

The qualitative analysis showed that 93% of the DLS generated OAR contours did not need corrections, except for the heart where 67% of the contours needed corrections. The majority of DLS generated CTVs needed corrections, whereas a minority were deemed not acceptable. Still, using the DLS-model for CTV and heart delineation is on average 14 minutes faster. An average DSC=0.91 and H95=9.8 mm were found for the left and right breasts, respectively. Likewise, and average DSC in the range [0.66, 0.76]mm and HD95 in the range [7.04, 12.05]mm were found for the lymph nodes.

CONCLUSION

The validation showed that the DLS generated OAR contours can be used clinically. Corrections were required to most of the DLS generated CTVs, and therefore warrants more attention before possibly implementing the DLS models clinically.

Development of a national deep learning-based auto-segmentation model for the heart on clinical delineations from the DBCG RT nation cohort

BACKGROUND

This study aimed at investigating the feasibility of developing a deep learning-based auto-segmentation model for the heart trained on clinical delineations.

MATERIAL AND METHODS

This study included two different datasets. The first dataset contained clinical heart delineations from the DBCG RT Nation study (1,561 patients). The second dataset was smaller (114 patients), but with corrected heart delineations. Before training the model on the clinical delineations an outlier-detection was performed, to remove cases with gross deviations from the delineation guideline. No outlier detection was performed for the dataset with corrected heart delineations. Both models were trained with a 3D full resolution nnUNet. The models were evaluated with the dice similarity coefficient (DSC), 95% Hausdorff distance (HD95) and Mean Surface Distance (MSD). The difference between the models were tested with the Mann-Whitney U-test. The balance of dataset quantity versus quality was investigated, by stepwise reducing the cohort size for the model trained on clinical delineations.

RESULTS

During the outlier-detection 137 patients were excluded from the clinical cohort due to non-compliance with delineation guidelines. The model trained on the curated clinical cohort performed with a median DSC of 0.96 (IQR 0.94-0.96), median HD95 of 4.00 mm (IQR 3.00 mm-6.00 mm) and a median MSD of 1.49 mm (IQR 1.12 mm-2.02 mm). The model trained on the dedicated and corrected cohort performed with a median DSC of 0.95 (IQR 0.93-0.96), median HD95 of 5.65 mm (IQR 3.37 mm-8.62 mm) and median MSD of 1.63 mm (IQR 1.35 mm-2.11 mm). The difference between the two models were found non-significant for all metrics (p > 0.05). Reduction of cohort size showed no significant difference for all metrics (p > 0.05). However, with the smallest cohort size, a few outlier structures were found.

CONCLUSIONS

This study demonstrated a deep learning-based auto-segmentation model trained on curated clinical delineations which performs on par with a model trained on dedicated delineations, making it easier to develop multi-institutional auto-segmentation models.

What Matters in Radiological Image Segmentation? Effect of Segmentation Errors on the Diagnostic Related Features

Segmentation is a crucial step in extracting the medical image features for clinical diagnosis. Though multiple metrics have been proposed to evaluate the segmentation performance, there is no clear study on how or to what extent the segmentation errors will affect the diagnostic related features used in clinical practice. Therefore, we proposed a segmentation robustness plot (SRP) to build the link between segmentation errors and clinical acceptance, where relative area under the curve (R-AUC) was designed to help clinicians to identify the robust diagnostic related image features. In experiments, we first selected representative radiological series from time series (cardiac first-pass perfusion) and spatial series (T2 weighted images on brain tumors) of magnetic resonance images, respectively. Then, dice similarity coefficient (DSC) and Hausdorff distance (HD), as the widely used evaluation metrics, were used to systematically control the degree of the segmentation errors. Finally, the differences between diagnostic related image features extracted from the ground truth and the derived segmentation were analyzed, using the statistical method large sample size T-test to calculate the corresponding p values. The results are denoted in the SRP, where the x-axis indicates the segmentation performance using the aforementioned evaluation metric, and the y-axis shows the severity of the corresponding feature changes, which are expressed in either the p values for a single case or the proportion of patients without significant change. The experimental results in SRP show that when DSC is above 0.95 and HD is below 3 mm, the segmentation errors will not change the features significantly in most cases. However, when segmentation gets worse, additional metrics are required for further analysis. In this way, the proposed SRP indicates the impact of the segmentation errors on the severity of the corresponding feature changes. By using SRP, one could easily define the acceptable segmentation errors in a challenge. Additionally, the R-AUC calculated from SRP provides an objective reference to help the selection of reliable features in image analysis.

Clinical Implementation and Evaluation of Auto-Segmentation Tools for Multi-Site Contouring in Radiotherapy

Background and purpose

Tools for auto-segmentation in radiotherapy are widely available, but guidelines for clinical implementation are missing. The goal was to develop a workflow for performance evaluation of three commercial auto-segmentation tools to select one candidate for clinical implementation.

Materials and Methods

One hundred patients with six treatment sites (brain, head-and-neck, thorax, abdomen, and pelvis) were included. Three sets of AI-based contours for organs-at-risk (OAR) generated by three software tools and manually drawn expert contours were blindly rated for contouring accuracy. The dice similarity coefficient (DSC), the Hausdorff distance, and a dose/volume evaluation based on the recalculation of the original treatment plan were assessed. Statistically significant differences were tested using the Kruskal-Wallis test and the post-hoc Dunn Test with Bonferroni correction.

Results

The mean DSC scores compared to expert contours for all OARs combined were 0.80 ± 0.10, 0.75 ± 0.10, and 0.74 ± 0.11 for the three software tools. Physicians' rating identified equivalent or superior performance of some AI-based contours in head (eye, lens, optic nerve, brain, chiasm), thorax (e.g., heart and lungs), and pelvis and abdomen (e.g., kidney, femoral head) compared to manual contours. For some OARs, the AI models provided results requiring only minor corrections. Bowel-bag and stomach were not fit for direct use. During the interdisciplinary discussion, the physicians' rating was considered the most relevant.

Conclusion

A comprehensive method for evaluation and clinical implementation of commercially available auto-segmentation software was developed. The in-depth analysis yielded clear instructions for clinical use within the radiotherapy department.

Determining The Role Of Radiation Oncologist Demographic Factors On Segmentation Quality: Insights From A Crowd-Sourced Challenge Using Bayesian Estimation

BACKGROUND

Medical image auto-segmentation is poised to revolutionize radiotherapy workflows. The quality of auto-segmentation training data, primarily derived from clinician observers, is of utmost importance. However, the factors influencing the quality of these clinician-derived segmentations have yet to be fully understood or quantified. Therefore, the purpose of this study was to determine the role of common observer demographic variables on quantitative segmentation performance.

METHODS

Organ at risk (OAR) and tumor volume segmentations provided by radiation oncologist observers from the Contouring Collaborative for Consensus in Radiation Oncology public dataset were utilized for this study. Segmentations were derived from five separate disease sites comprised of one patient case each: breast, sarcoma, head and neck (H&N), gynecologic (GYN), and gastrointestinal (GI). Segmentation quality was determined on a structure-by-structure basis by comparing the observer segmentations with an expert-derived consensus gold standard primarily using the Dice Similarity Coefficient (DSC); surface DSC was investigated as a secondary metric. Metrics were stratified into binary groups based on previously established structure-specific expert-derived interobserver variability (IOV) cutoffs. Generalized linear mixed-effects models using Markov chain Monte Carlo Bayesian estimation were used to investigate the association between demographic variables and the binarized segmentation quality for each disease site separately. Variables with a highest density interval excluding zero - loosely analogous to frequentist significance - were considered to substantially impact the outcome measure.

RESULTS

After filtering by practicing radiation oncologists, 574, 110, 452, 112, and 48 structure observations remained for the breast, sarcoma, H&N, GYN, and GI cases, respectively. The median percentage of observations that crossed the expert DSC IOV cutoff when stratified by structure type was 55% and 31% for OARs and tumor volumes, respectively. Bayesian regression analysis revealed tumor category had a substantial negative impact on binarized DSC for the breast (coefficient mean ± standard deviation: -0.97 ± 0.20), sarcoma (-1.04 ± 0.54), H&N (-1.00 ± 0.24), and GI (-2.95 ± 0.98) cases. There were no clear recurring relationships between segmentation quality and demographic variables across the cases, with most variables demonstrating large standard deviations and wide highest density intervals.

CONCLUSION

Our study highlights substantial uncertainty surrounding conventionally presumed factors influencing segmentation quality. Future studies should investigate additional demographic variables, more patients and imaging modalities, and alternative metrics of segmentation acceptability.

Investigating the feasibility of using Ethos generated treatment plans for head and neck cancer patients

The Varian Ethos treatment platform is designed to automatically create complex RT treatment plans, reducing both workload and operator variability in plan quality. The aim of this study is to evaluate the quality of Ethos-generated head and neck (H&N) treatment plans. Ethos plans were created for ten previous H&N patients and these were compared with the original clinical plans generated in Eclipse. Ethos automatically creates several plans with different field arrangements for each patient. All plans were compared quantitatively using: dose-volume metrics; dose conformity; dose heterogeneity and monitor units (MU). In addition, two H&N Oncologists assessed the clinical acceptability of the Ethos plans. Consultant 1 judged there to be at least three clinically acceptable Ethos plans for 9 out of 10 patients reviewed. Consultant 2 approved of at least two Ethos plans for 5 out of 5 patients reviewed. The Ethos plans' average dose metrics were comparable to the clinical plans. The average plan MU was similar for Eclipse and Ethos VMAT plans. The average plan MU for Ethos IMRT plans was larger with respect to all VMAT plans. The Ethos Treatment Planning system is capable of automatically creating good quality treatment plans for a range of H&N cancer patients.

Dosimetric comparison of autocontouring techniques for online adaptive proton therapy

Objective.Anatomical and daily set-up uncertainties impede high precision delivery of proton therapy. With online adaptation, the daily plan is reoptimized on an image taken shortly before the treatment, reducing these uncertainties and, hence, allowing a more accurate delivery. This reoptimization requires target and organs-at-risk (OAR) contours on the daily image, which need to be delineated automatically since manual contouring is too slow. Whereas multiple methods for autocontouring exist, none of them are fully accurate, which affects the daily dose. This work aims to quantify the magnitude of this dosimetric effect for four contouring techniques.Approach.Plans reoptimized on automatic contours are compared with plans reoptimized on manual contours. The methods include rigid and deformable registration (DIR), deep-learning based segmentation and patient-specific segmentation.Main results.It was found that independently of the contouring method, the dosimetric influence of usingautomaticOARcontoursis small (<5% prescribed dose in most cases), with DIR yielding the best results. Contrarily, the dosimetric effect of using theautomatic target contourwas larger (>5% prescribed dose in most cases), indicating that manual verification of that contour remains necessary. However, when compared to non-adaptive therapy, the dose differences caused by automatically contouring the target were small and target coverage was improved, especially for DIR.Significance.The results show that manual adjustment of OARs is rarely necessary and that several autocontouring techniques are directly usable. Contrarily, manual adjustment of the target is important. This allows prioritizing tasks during time-critical online adaptive proton therapy and therefore supports its further clinical implementation.

A clinical evaluation of the performance of five commercial artificial intelligence contouring systems for radiotherapy

Purpose/objectives

Auto-segmentation with artificial intelligence (AI) offers an opportunity to reduce inter- and intra-observer variability in contouring, to improve the quality of contours, as well as to reduce the time taken to conduct this manual task. In this work we benchmark the AI auto-segmentation contours produced by five commercial vendors against a common dataset.

Methods and materials

The organ at risk (OAR) contours generated by five commercial AI auto-segmentation solutions (Mirada (Mir), MVision (MV), Radformation (Rad), RayStation (Ray) and TheraPanacea (Ther)) were compared to manually-drawn expert contours from 20 breast, 20 head and neck, 20 lung and 20 prostate patients. Comparisons were made using geometric similarity metrics including volumetric and surface Dice similarity coefficient (vDSC and sDSC), Hausdorff distance (HD) and Added Path Length (APL). To assess the time saved, the time taken to manually draw the expert contours, as well as the time to correct the AI contours, were recorded.

Results

There are differences in the number of CT contours offered by each AI auto-segmentation solution at the time of the study (Mir 99; MV 143; Rad 83; Ray 67; Ther 86), with all offering contours of some lymph node levels as well as OARs. Averaged across all structures, the median vDSCs were good for all systems and compared favorably with existing literature: Mir 0.82; MV 0.88; Rad 0.86; Ray 0.87; Ther 0.88. All systems offer substantial time savings, ranging between: breast 14-20 mins; head and neck 74-93 mins; lung 20-26 mins; prostate 35-42 mins. The time saved, averaged across all structures, was similar for all systems: Mir 39.8 mins; MV 43.6 mins; Rad 36.6 min; Ray 43.2 mins; Ther 45.2 mins.

Conclusions

All five commercial AI auto-segmentation solutions evaluated in this work offer high quality contours in significantly reduced time compared to manual contouring, and could be used to render the radiotherapy workflow more efficient and standardized.