General and custom deep learning autosegmentation models for organs in head and neck, abdomen, and male pelvis

Accurate object localization facilitates automatic esophagus segmentation in deep learning

BACKGROUND

Currently, automatic esophagus segmentation remains a challenging task due to its small size, low contrast, and large shape variation. We aimed to improve the performance of esophagus segmentation in deep learning by applying a strategy that involves locating the object first and then performing the segmentation task.

METHODS

A total of 100 cases with thoracic computed tomography scans from two publicly available datasets were used in this study. A modified CenterNet, an object location network, was employed to locate the center of the esophagus for each slice. Subsequently, the 3D U-net and 2D U-net_coarse models were trained to segment the esophagus based on the predicted object center. A 2D U-net_fine model was trained based on the updated object center according to the 3D U-net model. The dice similarity coefficient and the 95% Hausdorff distance were used as quantitative evaluation indexes for the delineation performance. The characteristics of the automatically delineated esophageal contours by the 2D U-net and 3D U-net models were summarized. Additionally, the impact of the accuracy of object localization on the delineation performance was analyzed. Finally, the delineation performance in different segments of the esophagus was also summarized.

RESULTS

The mean dice coefficient of the 3D U-net, 2D U-net_coarse, and 2D U-net_fine models were 0.77, 0.81, and 0.82, respectively. The 95% Hausdorff distance for the above models was 6.55, 3.57, and 3.76, respectively. Compared with the 2D U-net, the 3D U-net has a lower incidence of delineating wrong objects and a higher incidence of missing objects. After using the fine object center, the average dice coefficient was improved by 5.5% in the cases with a dice coefficient less than 0.75, while that value was only 0.3% in the cases with a dice coefficient greater than 0.75. The dice coefficients were lower for the esophagus between the orifice of the inferior and the pulmonary bifurcation compared with the other regions.

CONCLUSION

The 3D U-net model tended to delineate fewer incorrect objects but also miss more objects. Two-stage strategy with accurate object location could enhance the robustness of the segmentation model and significantly improve the esophageal delineation performance, especially for cases with poor delineation results.

Deep learning and atlas-based models to streamline the segmentation workflow of total marrow and lymphoid irradiation

PURPOSE

To improve the workflow of total marrow and lymphoid irradiation (TMLI) by enhancing the delineation of organs at risk (OARs) and clinical target volume (CTV) using deep learning (DL) and atlas-based (AB) segmentation models.

MATERIALS AND METHODS

Ninety-five TMLI plans optimized in our institute were analyzed. Two commercial DL software were tested for segmenting 18 OARs. An AB model for lymph node CTV (CTV_LN) delineation was built using 20 TMLI patients. The AB model was evaluated on 20 independent patients, and a semiautomatic approach was tested by correcting the automatic contours. The generated OARs and CTV_LN contours were compared to manual contours in terms of topological agreement, dose statistics, and time workload. A clinical decision tree was developed to define a specific contouring strategy for each OAR.

RESULTS

The two DL models achieved a median [interquartile range] dice similarity coefficient (DSC) of 0.84 [0.71;0.93] and 0.85 [0.70;0.93] across the OARs. The absolute median Dmean difference between manual and the two DL models was 2.0 [0.7;6.6]% and 2.4 [0.9;7.1]%. The AB model achieved a median DSC of 0.70 [0.66;0.74] for CTV_LN delineation, increasing to 0.94 [0.94;0.95] after manual revision, with minimal Dmean differences. Since September 2022, our institution has implemented DL and AB models for all TMLI patients, reducing from 5 to 2 h the time required to complete the entire segmentation process.

CONCLUSION

DL models can streamline the TMLI contouring process of OARs. Manual revision is still necessary for lymph node delineation using AB models.

Artificial Intelligence-Based Radiotherapy Contouring and Planning to Improve Global Access to Cancer Care

PURPOSE

Increased automation has been identified as one approach to improving global cancer care. The Radiation Planning Assistant (RPA) is a web-based tool offering automated radiotherapy (RT) contouring and planning to low-resource clinics. In this study, the RPA workflow and clinical acceptability were assessed by physicians around the world.

METHODS

The RPA output for 75 cases was reviewed by at least three physicians; 31 radiation oncologists at 16 institutions in six countries on five continents reviewed RPA contours and plans for clinical acceptability using a 5-point Likert scale.

RESULTS

For cervical cancer, RPA plans using bony landmarks were scored as usable as-is in 81% (with minor edits 93%); using soft tissue contours, plans were scored as usable as-is in 79% (with minor edits 96%). For postmastectomy breast cancer, RPA plans were scored as usable as-is in 44% (with minor edits 91%). For whole-brain treatment, RPA plans were scored as usable as-is in 67% (with minor edits 99%). For head/neck cancer, the normal tissue autocontours were acceptable as-is in 89% (with minor edits 97%). The clinical target volumes (CTVs) were acceptable as-is in 40% (with minor edits 93%). The volumetric-modulated arc therapy (VMAT) plans were acceptable as-is in 87% (with minor edits 96%). For cervical cancer, the normal tissue autocontours were acceptable as-is in 92% (with minor edits 99%). The CTVs for cervical cancer were scored as acceptable as-is in 83% (with minor edits 92%). The VMAT plans for cervical cancer were acceptable as-is in 99% (with minor edits 100%).

CONCLUSION

The RPA, a web-based tool designed to improve access to high-quality RT in low-resource settings, has high rates of clinical acceptability by practicing clinicians around the world. It has significant potential for successful implementation in low-resource clinics.

Deep learning algorithm performance in contouring head and neck organs at risk: a systematic review and single-arm meta-analysis

PURPOSE

The contouring of organs at risk (OARs) in head and neck cancer radiation treatment planning is a crucial, yet repetitive and time-consuming process. Recent studies have applied deep learning (DL) algorithms to automatically contour head and neck OARs. This study aims to conduct a systematic review and meta-analysis to summarize and analyze the performance of DL algorithms in contouring head and neck OARs. The objective is to assess the advantages and limitations of DL algorithms in contour planning of head and neck OARs.

METHODS

This study conducted a literature search of Pubmed, Embase and Cochrane Library databases, to include studies related to DL contouring head and neck OARs, and the dice similarity coefficient (DSC) of four categories of OARs from the results of each study are selected as effect sizes for meta-analysis. Furthermore, this study conducted a subgroup analysis of OARs characterized by image modality and image type.

RESULTS

149 articles were retrieved, and 22 studies were included in the meta-analysis after excluding duplicate literature, primary screening, and re-screening. The combined effect sizes of DSC for brainstem, spinal cord, mandible, left eye, right eye, left optic nerve, right optic nerve, optic chiasm, left parotid, right parotid, left submandibular, and right submandibular are 0.87, 0.83, 0.92, 0.90, 0.90, 0.71, 0.74, 0.62, 0.85, 0.85, 0.82, and 0.82, respectively. For subgroup analysis, the combined effect sizes for segmentation of the brainstem, mandible, left optic nerve, and left parotid gland using CT and MRI images are 0.86/0.92, 0.92/0.90, 0.71/0.73, and 0.84/0.87, respectively. Pooled effect sizes using 2D and 3D images of the brainstem, mandible, left optic nerve, and left parotid gland for contouring are 0.88/0.87, 0.92/0.92, 0.75/0.71 and 0.87/0.85.

CONCLUSIONS

The use of automated contouring technology based on DL algorithms is an essential tool for contouring head and neck OARs, achieving high accuracy, reducing the workload of clinical radiation oncologists, and providing individualized, standardized, and refined treatment plans for implementing "precision radiotherapy". Improving DL performance requires the construction of high-quality data sets and enhancing algorithm optimization and innovation.

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.

Multi-organ segmentation of CT via convolutional neural network: impact of training setting and scanner manufacturer

Objective. Automated organ segmentation on CT images can enable the clinical use of advanced quantitative software devices, but model performance sensitivities must be understood before widespread adoption can occur. The goal of this study was to investigate performance differences between Convolutional Neural Networks (CNNs) trained to segment one (single-class) versus multiple (multi-class) organs, and between CNNs trained on scans from a single manufacturer versus multiple manufacturers.Methods. The multi-class CNN was trained on CT images obtained from 455 whole-body PET/CT scans (413 for training, 42 for testing) taken with Siemens, GE, and Phillips PET/CT scanners where 16 organs were segmented. The multi-class CNN was compared to 16 smaller single-class CNNs trained using the same data, but with segmentations of only one organ per model. In addition, CNNs trained on Siemens-only (N = 186) and GE-only (N = 219) scans (manufacturer-specific) were compared with CNNs trained on data from both Siemens and GE scanners (manufacturer-mixed). Segmentation performance was quantified using five performance metrics, including the Dice Similarity Coefficient (DSC).Results. The multi-class CNN performed well compared to previous studies, even in organs usually considered difficult auto-segmentation targets (e.g., pancreas, bowel). Segmentations from the multi-class CNN were significantly superior to those from smaller single-class CNNs in most organs, and the 16 single-class models took, on average, six times longer to segment all 16 organs compared to the single multi-class model. The manufacturer-mixed approach achieved minimally higher performance over the manufacturer-specific approach.Significance. A CNN trained on contours of multiple organs and CT data from multiple manufacturers yielded high-quality segmentations. Such a model is an essential enabler of image processing in a software device that quantifies and analyzes such data to determine a patient's treatment response. To date, this activity of whole organ segmentation has not been adopted due to the intense manual workload and time required.

Deep learning auto-segmentation on multi-sequence magnetic resonance images for upper abdominal organs

Introduction

Multi-sequence multi-parameter MRIs are often used to define targets and/or organs at risk (OAR) in radiation therapy (RT) planning. Deep learning has so far focused on developing auto-segmentation models based on a single MRI sequence. The purpose of this work is to develop a multi-sequence deep learning based auto-segmentation (mS-DLAS) based on multi-sequence abdominal MRIs.

Materials and methods

Using a previously developed 3DResUnet network, a mS-DLAS model using 4 T1 and T2 weighted MRI acquired during routine RT simulation for 71 cases with abdominal tumors was trained and tested. Strategies including data pre-processing, Z-normalization approach, and data augmentation were employed. Additional 2 sequence specific T1 weighted (T1-M) and T2 weighted (T2-M) models were trained to evaluate performance of sequence-specific DLAS. Performance of all models was quantitatively evaluated using 6 surface and volumetric accuracy metrics.

Results

The developed DLAS models were able to generate reasonable contours of 12 upper abdomen organs within 21 seconds for each testing case. The 3D average values of dice similarity coefficient (DSC), mean distance to agreement (MDA mm), 95 percentile Hausdorff distance (HD95% mm), percent volume difference (PVD), surface DSC (sDSC), and relative added path length (rAPL mm/cc) over all organs were 0.87, 1.79, 7.43, -8.95, 0.82, and 12.25, respectively, for mS-DLAS model. Collectively, 71% of the auto-segmented contours by the three models had relatively high quality. Additionally, the obtained mS-DLAS successfully segmented 9 out of 16 MRI sequences that were not used in the model training.

Conclusion

We have developed an MRI-based mS-DLAS model for auto-segmenting of upper abdominal organs on MRI. Multi-sequence segmentation is desirable in routine clinical practice of RT for accurate organ and target delineation, particularly for abdominal tumors. Our work will act as a stepping stone for acquiring fast and accurate segmentation on multi-contrast MRI and make way for MR only guided radiation therapy.

Computational approaches for the reconstruction of optic nerve fibers along the visual pathway from medical images: a comprehensive review

Optic never fibers in the visual pathway play significant roles in vision formation. Damages of optic nerve fibers are biomarkers for the diagnosis of various ophthalmological and neurological diseases; also, there is a need to prevent the optic nerve fibers from getting damaged in neurosurgery and radiation therapy. Reconstruction of optic nerve fibers from medical images can facilitate all these clinical applications. Although many computational methods are developed for the reconstruction of optic nerve fibers, a comprehensive review of these methods is still lacking. This paper described both the two strategies for optic nerve fiber reconstruction applied in existing studies, i.e., image segmentation and fiber tracking. In comparison to image segmentation, fiber tracking can delineate more detailed structures of optic nerve fibers. For each strategy, both conventional and AI-based approaches were introduced, and the latter usually demonstrates better performance than the former. From the review, we concluded that AI-based methods are the trend for optic nerve fiber reconstruction and some new techniques like generative AI can help address the current challenges in optic nerve fiber reconstruction.

Validation of clinical acceptability of deep-learning-based automated segmentation of organs-at-risk for head-and-neck radiotherapy treatment planning

Introduction

Organ-at-risk segmentation for head and neck cancer radiation therapy is a complex and time-consuming process (requiring up to 42 individual structure, and may delay start of treatment or even limit access to function-preserving care. Feasibility of using a deep learning (DL) based autosegmentation model to reduce contouring time without compromising contour accuracy is assessed through a blinded randomized trial of radiation oncologists (ROs) using retrospective, de-identified patient data.

Methods

Two head and neck expert ROs used dedicated time to create gold standard (GS) contours on computed tomography (CT) images. 445 CTs were used to train a custom 3D U-Net DL model covering 42 organs-at-risk, with an additional 20 CTs were held out for the randomized trial. For each held-out patient dataset, one of the eight participant ROs was randomly allocated to review and revise the contours produced by the DL model, while another reviewed contours produced by a medical dosimetry assistant (MDA), both blinded to their origin. Time required for MDAs and ROs to contour was recorded, and the unrevised DL contours, as well as the RO-revised contours by the MDAs and DL model were compared to the GS for that patient.

Results

Mean time for initial MDA contouring was 2.3 hours (range 1.6-3.8 hours) and RO-revision took 1.1 hours (range, 0.4-4.4 hours), compared to 0.7 hours (range 0.1-2.0 hours) for the RO-revisions to DL contours. Total time reduced by 76% (95%-Confidence Interval: 65%-88%) and RO-revision time reduced by 35% (95%-CI,-39%-91%). All geometric and dosimetric metrics computed, agreement with GS was equivalent or significantly greater (p<0.05) for RO-revised DL contours compared to the RO-revised MDA contours, including volumetric Dice similarity coefficient (VDSC), surface DSC, added path length, and the 95%-Hausdorff distance. 32 OARs (76%) had mean VDSC greater than 0.8 for the RO-revised DL contours, compared to 20 (48%) for RO-revised MDA contours, and 34 (81%) for the unrevised DL OARs.

Conclusion

DL autosegmentation demonstrated significant time-savings for organ-at-risk contouring while improving agreement with the institutional GS, indicating comparable accuracy of DL model. Integration into the clinical practice with a prospective evaluation is currently underway.

Deep learning based automatic contour refinement for inaccurate auto-segmentation in MR-guided adaptive radiotherapy

Objective.Fast and accurate auto-segmentation is essential for magnetic resonance-guided adaptive radiation therapy (MRgART). Deep learning auto-segmentation (DLAS) is not always clinically acceptable, particularly for complex abdominal organs. We previously reported an automatic contour refinement (ACR) solution of using an active contour model (ACM) to partially correct the DLAS contours. This study aims to develop a DL-based ACR model to work in conjunction with ACM-ACR to further improve the contour accuracy.Approach.The DL-ACR model was trained and tested using bowel contours created by an in-house DLAS system from 160 MR sets (76 from MR-simulation and 84 from MR-Linac). The contours were classified into acceptable, minor-error and major-error groups using two approaches of contour quality classification (CQC), based on the AAPM TG-132 recommendation and an in-house classification model, respectively. For the major-error group, DL-ACR was applied subsequently after ACM-ACR to further refine the contours. For the minor-error group, contours were directly corrected by DL-ACR without applying an initial ACM-ACR. The ACR workflow was performed separately for the two CQC methods and was evaluated using contours from 25 image sets as independent testing data.Main results.The best ACR performance was observed in the MR-simulation testing set using CQC by TG-132: (1) for the major-error group, 44% (177/401) were improved to minor-error group and 5% (22/401) became acceptable by applying ACM-ACR; among these 177 contours that shifted from major-error to minor-error with ACM-ACR, DL-ACR further refined 49% (87/177) to acceptable; and overall, 36% (145/401) were improved to minor-error contours, and 30% (119/401) became acceptable after sequentially applying ACM-ACR and DL-ACR; (2) for the minor-error group, 43% (320/750) were improved to acceptable contours using DL-ACR.Significance.The obtained ACR workflow substantially improves the accuracy of DLAS bowel contours, minimizing the manual editing time and accelerating the segmentation process of MRgART.

Anatomical evaluation of deep-learning synthetic computed tomography images generated from male pelvis cone-beam computed tomography

Background and purpose

To improve cone-beam computed tomography (CBCT), deep-learning (DL)-models are being explored to generate synthetic CTs (sCT). The sCT evaluation is mainly focused on image quality and CT number accuracy. However, correct representation of daily anatomy of the CBCT is also important for sCTs in adaptive radiotherapy. The aim of this study was to emphasize the importance of anatomical correctness by quantitatively assessing sCT scans generated from CBCT scans using different paired and unpaired dl-models.

Materials and methods

Planning CTs (pCT) and CBCTs of 56 prostate cancer patients were included to generate sCTs. Three different dl-models, Dual-UNet, Single-UNet and Cycle-consistent Generative Adversarial Network (CycleGAN), were evaluated on image quality and anatomical correctness. The image quality was assessed using image metrics, such as Mean Absolute Error (MAE). The anatomical correctness between sCT and CBCT was quantified using organs-at-risk volumes and average surface distances (ASD).

Results

MAE was 24 Hounsfield Unit (HU) [range:19-30 HU] for Dual-UNet, 40 HU [range:34-56 HU] for Single-UNet and 41HU [range:37-46 HU] for CycleGAN. Bladder ASD was 4.5 mm [range:1.6-12.3 mm] for Dual-UNet, 0.7 mm [range:0.4-1.2 mm] for Single-UNet and 0.9 mm [range:0.4-1.1 mm] CycleGAN.

Conclusions

Although Dual-UNet performed best in standard image quality measures, such as MAE, the contour based anatomical feature comparison with the CBCT showed that Dual-UNet performed worst on anatomical comparison. This emphasizes the importance of adding anatomy based evaluation of sCTs generated by dl-models. For applications in the pelvic area, direct anatomical comparison with the CBCT may provide a useful method to assess the clinical applicability of dl-based sCT generation methods.

An uncertainty-aware deep learning architecture with outlier mitigation for prostate gland segmentation in radiotherapy treatment planning

PURPOSE

Task automation is essential for efficient and consistent image segmentation in radiation oncology. We report on a deep learning architecture, comprising a U-Net and a variational autoencoder (VAE) for automatic contouring of the prostate gland incorporating interobserver variation for radiotherapy treatment planning. The U-Net/VAE generates an ensemble set of segmentations for each image CT slice. A novel outlier mitigation (OM) technique was implemented to enhance the model segmentation accuracy.

METHODS

The primary source dataset (source_prim) consisted of 19 200 CT slices (from 300 patient planning CT image datasets) with manually contoured prostate glands. A smaller secondary source dataset (source_sec) comprised 640 CT slices (from 10 patient CT datasets), where prostate glands were segmented by 5 independent physicians on each dataset to account for interobserver variability. Data augmentation via random rotation (<5 degrees), cropping, and horizontal flipping was applied to each dataset to increase sample size by a factor of 100. A probabilistic hierarchical U-Net with VAE was implemented and pretrained using the augmented source_prim dataset for 30 epochs. Model parameters of the U-Net/VAE were fine-tuned using the augmented source_sec dataset for 100 epochs. After the first round of training, outlier contours in the training dataset were automatically detected and replaced by the most accurate contours (based on Dice similarity coefficient, DSC) generated by the model. The U-Net/OM-VAE was retrained using the revised training dataset. Metrics for comparison included DSC, Hausdorff distance (HD, mm), normalized cross-correlation (NCC) coefficient, and center-of-mass (COM) distance (mm).

RESULTS

Results for U-Net/OM-VAE with outliers replaced in the training dataset versus U-Net/VAE without OM were as follows: DSC = 0.82 ± 0.01 versus 0.80 ± 0.02 (p = 0.019), HD = 9.18 ± 1.22 versus 10.18 ± 1.35 mm (p = 0.043), NCC = 0.59 ± 0.07 versus 0.62 ± 0.06, and COM = 3.36 ± 0.81 versus 4.77 ± 0.96 mm over the average of 15 contours. For the average of 15 highest accuracy contours, values were as follows: DSC = 0.90 ± 0.02 versus 0.85 ± 0.02, HD = 5.47 ± 0.02 versus 7.54 ± 1.36 mm, and COM = 1.03 ± 0.58 versus 1.46 ± 0.68 mm (p < 0.03 for all metrics). Results for the U-Net/OM-VAE with outliers removed were as follows: DSC = 0.78 ± 0.01, HD = 10.65 ± 1.95 mm, NCC = 0.46 ± 0.10, COM = 4.17 ± 0.79 mm for the average of 15 contours, and DSC = 0.88 ± 0.02, HD = 7.00 ± 1.17 mm, COM = 1.58 ± 0.63 mm for the average of 15 highest accuracy contours. All metrics for U-Net/VAE trained on the source_prim and source_sec datasets via pretraining, followed by fine-tuning, show statistically significant improvement over that trained on the source_sec dataset only. Finally, all metrics for U-Net/VAE with or without OM showed statistically significant improvement over those for the standard U-Net.

CONCLUSIONS

A VAE combined with a hierarchical U-Net and an OM strategy (U-Net/OM-VAE) demonstrates promise toward capturing interobserver variability and produces accurate prostate auto-contours for radiotherapy planning. The availability of multiple contours for each CT slice enables clinicians to determine trade-offs in selecting the "best fitting" contour on each CT slice. Mitigation of outlier contours in the training dataset improves prediction accuracy, but one must be wary of reduction in variability in the training dataset.

Comparison of atlas-based and deep learning methods for organs at risk delineation on head-and-neck CT images using an automated treatment planning system

BACKGROUND AND PURPOSE

To investigate the performance of head-and-neck (HN) organs-at-risk (OAR) automatic segmentation (AS) using four atlas-based (ABAS) and two deep learning (DL) solutions.

MATERIAL AND METHODS

All patients underwent iodine contrast-enhanced planning CT. Fourteen OAR were manually delineated. DL.1 and DL.2 solutions were trained with 63 mono-centric patients and > 1000 multi-centric patients, respectively. Ten and 15 patients with varied anatomies were selected for the atlas library and for testing, respectively. The evaluation was based on geometric indices (DICE coefficient and 95th percentile-Hausdorff Distance (HD_95%)), time needed for manual corrections and clinical dosimetric endpoints obtained using automated treatment planning.

RESULTS

Both DICE and HD_95% results indicated that DL algorithms generally performed better compared with ABAS algorithms for automatic segmentation of HN OAR. However, the hybrid-ABAS (ABAS.3) algorithm sometimes provided the highest agreement to the reference contours compared with the 2 DL. Compared with DL.2 and ABAS.3, DL.1 contours were the fastest to correct. For the 3 solutions, the differences in dose distributions obtained using AS contours and AS + manually corrected contours were not statistically significant. High dose differences could be observed when OAR contours were at short distances to the targets. However, this was not always interrelated.

CONCLUSION

DL methods generally showed higher delineation accuracy compared with ABAS methods for AS segmentation of HN OAR. Most ABAS contours had high conformity to the reference but were more time consuming than DL algorithms, especially when considering the computing time and the time spent on manual corrections.

A Prior Knowledge-Guided, Deep Learning-Based Semiautomatic Segmentation for Complex Anatomy on Magnetic Resonance Imaging

PURPOSE

Despite recent substantial improvement in autosegmentation using deep learning (DL) methods, labor-intensive and time-consuming slice-by-slice manual editing is often needed, particularly for complex anatomy (eg, abdominal organs). This work aimed to develop a fast, prior knowledge-guided DL semiautomatic segmentation (DL-SAS) method for complex structures on abdominal magnetic resonance imaging (MRI) scans.

METHODS AND MATERIALS

A novel application using contours on an adjacent slice as a prior knowledge informant in a 2-dimensional UNet DL model to guide autosegmentation for a subsequent slice was implemented for DL-SAS. A generalized, instead of organ-specific, DL-SAS model was trained and tested for abdominal organs on T2-weighted MRI scans collected from 75 patients (65 for training and 10 for testing). The DL-SAS model performance was compared with 3 common autocontouring methods (linear interpolation, rigid propagation, and a full 3-dimensional DL autosegmentation model trained with the same training data set) based on various quantitative metrics including the Dice similarity coefficient (DSC) and ratio of acceptable slices (ROA) using paired t tests.

RESULTS

For the 10 testing cases, the DL-SAS model performed best with the slice interval (SI) of 1, resulting in an average DSC of 0.93 ± 0.02, 0.92 ± 0.02, 0.91 ± 0.02, 0.88 ± 0.03, and 0.87 ± 0.02 for the large bowel, stomach, small bowel, duodenum, and pancreas, respectively. The performance decreased with increased SIs from the guidance slice. The DL-SAS method performed significantly better (P < .05) than the other 3 methods. The ROA values were in the range of 48% to 66% for all the organs with an SI of 1 for DL-SAS, higher than those for linear interpolation (31%-57% for an SI of 1) and DL auto-segmentation (16%-51%).

CONCLUSIONS

The developed DL-SAS model segmented complex abdominal structures on MRI with high accuracy and efficiency and may be implemented as an interactive manual contouring tool or a contour editing tool in conjunction with a full autosegmentation process, facilitating fast and accurate segmentation for MRI-guided online adaptive radiation therapy.

Automatic contour refinement for deep learning auto-segmentation of complex organs in MRI-guided adaptive radiotherapy

Purpose

Fast and accurate auto-segmentation on daily images is essential for magnetic resonance imaging (MRI)–guided adaptive radiation therapy (ART). However, the state-of-the-art auto-segmentation based on deep learning still has limited success, particularly for complex structures in the abdomen. This study aimed to develop an automatic contour refinement (ACR) process to quickly correct for unacceptable auto-segmented contours.

Methods and Materials

An improved level set–based active contour model (ACM) was implemented for the ACR process and was tested on the deep learning–based auto-segmentation of 80 abdominal MRI sets along with their ground truth contours. The performance of the ACR process was evaluated using 4 contour accuracy metrics: the Dice similarity coefficient (DSC), mean distance to agreement (MDA), surface DSC, and added path length (APL) on the auto-segmented contours of the small bowel, large bowel, combined bowels, pancreas, duodenum, and stomach.

Results

A portion (3%-39%) of the corrected contours became practically acceptable per the American Association of Physicists in Medicine Task Group 132 (TG-132) recommendation (DSC >0.8 and MDA <3 mm). The best correction performance was seen in the combined bowels, where for the contours with major errors (initial DSC <0.5 or MDA >8 mm), the mean DSC increased from 0.34 to 0.59, the mean MDA decreased from 7.02 mm to 5.23 mm, and the APL reduced by almost 20 mm, whereas for the contours with minor errors, the mean DSC increased from 0.72 to 0.79, the mean MDA decreased from 3.35 mm to 3.29 mm, and more than one-third (39%) of the ACR contours became clinically acceptable. The execution time for the ACR process on one subregion was less than 2 seconds using an NVIDIA GTX 1060 GPU.

Conclusions

The ACR process implemented based on the ACM was able to quickly correct for some inaccurate contours produced from MRI-based deep learning auto-segmentation of complex abdominal anatomy. The ACR method may be integrated into the auto-segmentation step to accelerate the process of MRI-guided ART.

Optimising a 3D convolutional neural network for head and neck computed tomography segmentation with limited training data

•

Convolutional neural networks (CNNs) are used for auto-segmentation in radiotherapy.

•

However, CNNs rely on large, high-quality datasets: a scarcity in radiotherapy.

•

We develop a CNN model, trained with limited data, for accurate segmentation.

•

Multiple experiments were performed to optimise key features of our custom model.

•

Our model is competitive with state-of-the-art methods on a public dataset.

Background and purpose

Convolutional neural networks (CNNs) are increasingly used to automate segmentation for radiotherapy planning, where accurate segmentation of organs-at-risk (OARs) is crucial. Training CNNs often requires large amounts of data. However, large, high quality datasets are scarce. The aim of this study was to develop a CNN capable of accurate head and neck (HN) 3D auto-segmentation of planning CT scans using a small training dataset (34 CTs).

Materials and Method

Elements of our custom CNN architecture were varied to optimise segmentation performance. We tested and evaluated the impact of: using multiple contrast channels for the CT scan input at specific soft tissue and bony anatomy windows, resize vs. transpose convolutions, and loss functions based on overlap metrics and cross-entropy in different combinations. Model segmentation performance was compared with the inter-observer deviation of two doctors’ gold standard segmentations using the 95th percentile Hausdorff distance and mean distance-to-agreement (mDTA). The best performing configuration was further validated on a popular public dataset to compare with state-of-the-art (SOTA) auto-segmentation methods.

Results

Our best performing CNN configuration was competitive with current SOTA methods when evaluated on the public dataset with mDTA of (0.81±0.31) mm for the brainstem, (0.20±0.08) mm for the mandible, (0.77±0.14) mm for the left parotid and (0.81±0.28) mm for the right parotid.

Conclusions

Through careful tuning and customisation we trained a 3D CNN with a small dataset to produce segmentations of HN OARs with an accuracy that is comparable with inter-clinician deviations. Our proposed model performed competitively with current SOTA methods.