Deep Convolution Neural Network (DCNN) Multiplane Approach to Synthetic CT Generation From MR images—Application in Brain Proton Therapy

Feasibility of Monte Carlo dropout-based uncertainty maps to evaluate deep learning-based synthetic CTs for adaptive proton therapy

BACKGROUND

Deep learning has shown promising results to generate MRI-based synthetic CTs and to enable accurate proton dose calculations on MRIs. For clinical implementation of synthetic CTs, quality assurance tools that verify their quality and reliability are required but still lacking.

PURPOSE

This study aims to evaluate the predictive value of uncertainty maps generated with Monte Carlo dropout (MCD) for verifying proton dose calculations on deep-learning-based synthetic CTs (sCTs) derived from MRIs in online adaptive proton therapy.

METHODS

Two deep-learning models (DCNN and cycleGAN) were trained for CT image synthesis using 101 paired CT-MR images. sCT images were generated using MCD for each model by performing 10 inferences with activated dropout layers. The final sCT was obtained by averaging the inferred sCTs, while the uncertainty map was obtained from the HU variance corresponding to each voxel of 10 sCTs. The resulting uncertainty maps were compared to the observed HU-, range-, WET-, and dose-error maps between the sCT and planning CT. For range and WET errors, the generated uncertainty maps were projected along the 90-degree angle. To evaluate the dose distribution, a mask based on the 5%-isodose curve was applied to only include voxels along the beam paths. Pearson's correlation coefficients were calculated to determine the correlation between the uncertainty maps and HUs, range, WET, and dose errors. To evaluate the dosimetric accuracy of synthetic CTs, clinical proton treatment plans were recalculated and compared to the pCTs RESULTS: Evaluation of the correlation showed an average of r = 0.92 ± 0.03 and r = 0.92 ± 0.03 for errors between uncertainty-HU, r = 0.66 ± 0.09 and r = 0.62 ± 0.06 between uncertainty-range, r = 0.64 ± 0.06 and r = 0.58 ± 0.07 between uncertainty-WET, and r = 0.65 ± 0.09 and r = 0.67 ± 0.07 between uncertainty and dose difference for DCNN and cycleGAN model, respectively. Dosimetric comparison for target volumes showed an average 3%/3 mm gamma pass rate of 99.76 ± 0.43 (DCNN) and 99.10 ± 1.27 (cycleGAN).

CONCLUSION

The observed correlations between uncertainty maps and the various metrics (HU, range, WET, and dose errors) demonstrated the potential of MCD-based uncertainty maps as a reliable QA tool to evaluate the accuracy of deep learning-based sCTs.

SC-GAN: Structure-completion generative adversarial network for synthetic CT generation from MR images with truncated anatomy

Creating synthetic CT (sCT) from magnetic resonance (MR) images enables MR-based treatment planning in radiation therapy. However, the MR images used for MR-guided adaptive planning are often truncated in the boundary regions due to the limited field of view and the need for sequence optimization. Consequently, the sCT generated from these truncated MR images lacks complete anatomic information, leading to dose calculation error for MR-based adaptive planning. We propose a novel structure-completion generative adversarial network (SC-GAN) to generate sCT with full anatomic details from the truncated MR images. To enable anatomy compensation, we expand input channels of the CT generator by including a body mask and introduce a truncation loss between sCT and real CT. The body mask for each patient was automatically created from the simulation CT scans and transformed to daily MR images by rigid registration as another input for our SC-GAN in addition to the MR images. The truncation loss was constructed by implementing either an auto-segmentor or an edge detector to penalize the difference in body outlines between sCT and real CT. The experimental results show that our SC-GAN achieved much improved accuracy of sCT generation in both truncated and untruncated regions compared to the original cycleGAN and conditional GAN methods.

A systematic literature review: deep learning techniques for synthetic medical image generation and their applications in radiotherapy

The aim of this systematic review is to determine whether Deep Learning (DL) algorithms can provide a clinically feasible alternative to classic algorithms for synthetic Computer Tomography (sCT). The following categories are presented in this study: ∙ MR-based treatment planning and synthetic CT generation techniques. ∙ Generation of synthetic CT images based on Cone Beam CT images. ∙ Low-dose CT to High-dose CT generation. ∙ Attenuation correction for PET images. To perform appropriate database searches, we reviewed journal articles published between January 2018 and June 2023. Current methodology, study strategies, and results with relevant clinical applications were analyzed as we outlined the state-of-the-art of deep learning based approaches to inter-modality and intra-modality image synthesis. This was accomplished by contrasting the provided methodologies with traditional research approaches. The key contributions of each category were highlighted, specific challenges were identified, and accomplishments were summarized. As a final step, the statistics of all the cited works from various aspects were analyzed, which revealed that DL-based sCTs have achieved considerable popularity, while also showing the potential of this technology. In order to assess the clinical readiness of the presented methods, we examined the current status of DL-based sCT generation.

Optimized models and deep learning methods for drug response prediction in cancer treatments: a review

Recent advancements in deep learning (DL) have played a crucial role in aiding experts to develop personalized healthcare services, particularly in drug response prediction (DRP) for cancer patients. The DL's techniques contribution to this field is significant, and they have proven indispensable in the medical field. This review aims to analyze the diverse effectiveness of various DL models in making these predictions, drawing on research published from 2017 to 2023. We utilized the VOS-Viewer 1.6.18 software to create a word cloud from the titles and abstracts of the selected studies. This study offers insights into the focus areas within DL models used for drug response. The word cloud revealed a strong link between certain keywords and grouped themes, highlighting terms such as deep learning, machine learning, precision medicine, precision oncology, drug response prediction, and personalized medicine. In order to achieve an advance in DRP using DL, the researchers need to work on enhancing the models' generalizability and interoperability. It is also crucial to develop models that not only accurately represent various architectures but also simplify these architectures, balancing the complexity with the predictive capabilities. In the future, researchers should try to combine methods that make DL models easier to understand; this will make DRP reviews more open and help doctors trust the decisions made by DL models in cancer DRP.

Channel-wise attention enhanced and structural similarity constrained cycleGAN for effective synthetic CT generation from head and neck MRI images

BACKGROUND

Magnetic resonance imaging (MRI) plays an increasingly important role in radiotherapy, enhancing the accuracy of target and organs at risk delineation, but the absence of electron density information limits its further clinical application. Therefore, the aim of this study is to develop and evaluate a novel unsupervised network (cycleSimulationGAN) for unpaired MR-to-CT synthesis.

METHODS

The proposed cycleSimulationGAN in this work integrates contour consistency loss function and channel-wise attention mechanism to synthesize high-quality CT-like images. Specially, the proposed cycleSimulationGAN constrains the structural similarity between the synthetic and input images for better structural retention characteristics. Additionally, we propose to equip a novel channel-wise attention mechanism based on the traditional generator of GAN to enhance the feature representation capability of deep network and extract more effective features. The mean absolute error (MAE) of Hounsfield Units (HU), peak signal-to-noise ratio (PSNR), root-mean-square error (RMSE) and structural similarity index (SSIM) were calculated between synthetic CT (sCT) and ground truth (GT) CT images to quantify the overall sCT performance.

RESULTS

One hundred and sixty nasopharyngeal carcinoma (NPC) patients who underwent volumetric-modulated arc radiotherapy (VMAT) were enrolled in this study. The generated sCT of our method were more consistent with the GT compared with other methods in terms of visual inspection. The average MAE, RMSE, PSNR, and SSIM calculated over twenty patients were 61.88 ± 1.42, 116.85 ± 3.42, 36.23 ± 0.52 and 0.985 ± 0.002 for the proposed method. The four image quality assessment metrics were significantly improved by our approach compared to conventional cycleGAN, the proposed cycleSimulationGAN produces significantly better synthetic results except for SSIM in bone.

CONCLUSIONS

We developed a novel cycleSimulationGAN model that can effectively create sCT images, making them comparable to GT images, which could potentially benefit the MRI-based treatment planning.

Deep learning based synthesis of MRI, CT and PET: Review and analysis

Medical image synthesis represents a critical area of research in clinical decision-making, aiming to overcome the challenges associated with acquiring multiple image modalities for an accurate clinical workflow. This approach proves beneficial in estimating an image of a desired modality from a given source modality among the most common medical imaging contrasts, such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET). However, translating between two image modalities presents difficulties due to the complex and non-linear domain mappings. Deep learning-based generative modelling has exhibited superior performance in synthetic image contrast applications compared to conventional image synthesis methods. This survey comprehensively reviews deep learning-based medical imaging translation from 2018 to 2023 on pseudo-CT, synthetic MR, and synthetic PET. We provide an overview of synthetic contrasts in medical imaging and the most frequently employed deep learning networks for medical image synthesis. Additionally, we conduct a detailed analysis of each synthesis method, focusing on their diverse model designs based on input domains and network architectures. We also analyse novel network architectures, ranging from conventional CNNs to the recent Transformer and Diffusion models. This analysis includes comparing loss functions, available datasets and anatomical regions, and image quality assessments and performance in other downstream tasks. Finally, we discuss the challenges and identify solutions within the literature, suggesting possible future directions. We hope that the insights offered in this survey paper will serve as a valuable roadmap for researchers in the field of medical image synthesis.

Uncertainty-aware MR-based CT synthesis for robust proton therapy planning of brain tumour

BACKGROUND AND PURPOSE

Deep learning techniques excel in MR-based CT synthesis, but missing uncertainty prediction limits its clinical use in proton therapy. We developed an uncertainty-aware framework and evaluated its efficiency in robust proton planning.

MATERIALS AND METHODS

A conditional generative-adversarial network was trained on 64 brain tumour patients with paired MR-CT images to generate synthetic CTs (sCT) from combined T1-T2 MRs of three orthogonal planes. A Bayesian neural network predicts Laplacian distributions for all voxels with parameters (μ, b). A robust proton plan was optimized using three sCTs of μ and μ±b. The dosimetric differences between the plan from sCT (sPlan) and the recalculated plan (rPlan) on planning CT (pCT) were quantified for each patient. The uncertainty-aware robust plan was compared to conventional robust (global ± 3 %) and non-robust plans.

RESULTS

In 8-fold cross-validation, sCT-pCT image differences (Mean-Absolute-Error) were 80.84 ± 9.84HU (body), 35.78 ± 6.07HU (soft tissues) and 221.88 ± 31.69HU (bones), with Dice scores of 90.33 ± 2.43 %, 95.13 ± 0.80 %, and 85.53 ± 4.16 %, respectively. The uncertainty distribution positively correlated with absolute prediction error (Correlation Coefficient: 0.62 ± 0.01). The uncertainty-conditioned robust optimisation improved the rPlan-sPlan agreement, e.g., D95 absolute difference (CTV) was 1.10 ± 1.24 % compared to conventional (1.64 ± 2.71 %) and non-robust (2.08 ± 2.96 %) optimisation. This trend was consistent across all target and organs-at-risk indexes.

CONCLUSION

The enhanced framework incorporates 3D uncertainty prediction and generates high-quality sCTs from MR images. The framework also facilitates conditioned robust optimisation, bolstering proton plan robustness against network prediction errors. The innovative feature of uncertainty visualisation and robust analyses contribute to evaluating sCT clinical utility for individual patients.

A study of Bayesian deep network uncertainty and its application to synthetic CT generation for MR-only radiotherapy treatment planning

BACKGROUND

The use of synthetic computed tomography (CT) for radiotherapy treatment planning has received considerable attention because of the absence of ionizing radiation and close spatial correspondence to source magnetic resonance (MR) images, which have excellent tissue contrast. However, in an MR-only environment, little effort has been made to examine the quality of synthetic CT images without using the original CT images.

PURPOSE

To estimate synthetic CT quality without referring to original CT images, this study established the relationship between synthetic CT uncertainty and Bayesian uncertainty, and proposed a new Bayesian deep network for generating synthetic CT images and estimating synthetic CT uncertainty for MR-only radiotherapy treatment planning.

METHODS AND MATERIALS

A novel deep Bayesian network was formulated using probabilistic network weights. Two mathematical expressions were proposed to quantify the Bayesian uncertainty of the network and synthetic CT uncertainty, which was closely related to the mean absolute error (MAE) in Hounsfield Unit (HU) of synthetic CT. These uncertainties were examined to demonstrate the accuracy of representing the synthetic CT uncertainty using a Bayesian counterpart. We developed a hybrid Bayesian architecture and a new data normalization scheme, enabling the Bayesian network to generate both accurate synthetic CT and reliable uncertainty information when probabilistic weights were applied. The proposed method was evaluated in 59 patients (13/12/32/2 for training/validation/testing/uncertainty visualization) diagnosed with prostate cancer, who underwent same-day pelvic CT- and MR-acquisitions. To assess the relationship between Bayesian and synthetic CT uncertainties, linear and non-linear correlation coefficients were calculated on per-voxel, per-tissue, and per-patient bases. For accessing the accuracy of the CT number and dosimetric accuracy, the proposed method was compared with a commercially available atlas-based method (MRCAT) and a U-Net conditional-generative adversarial network (UcGAN).

RESULTS

The proposed model exhibited 44.33 MAE, outperforming UcGAN 52.51 and MRCAT 54.87. The gamma rate (2%/2 mm dose difference/distance to agreement) of the proposed model was 98.68%, comparable to that of UcGAN (98.60%) and MRCAT (98.56%). The per-patient and per-tissue linear correlation coefficients between the Bayesian and synthetic CT uncertainties ranged from 0.53 to 0.83, implying a moderate to strong linear correlation. Per-voxel correlation coefficients varied from -0.13 to 0.67 depending on the regions-of-interest evaluated, indicating tissue-dependent correlation. The R2 value for estimating MAE solely using Bayesian uncertainty was 0.98, suggesting that the uncertainty of the proposed model was an ideal candidate for predicting synthetic CT error, without referring to the original CT.

CONCLUSION

This study established a relationship between the Bayesian model uncertainty and synthetic CT uncertainty. A novel Bayesian deep network was proposed to generate a synthetic CT and estimate its uncertainty. Various metrics were used to thoroughly examine the relationship between the uncertainties of the proposed Bayesian model and the generated synthetic CT. Compared with existing approaches, the proposed model showed comparable CT number and dosimetric accuracies. The experiments showed that the proposed Bayesian model was capable of producing accurate synthetic CT, and was an effective indicator of the uncertainty and error associated with synthetic CT in MR-only workflows.

Mitigating misalignment in MRI-to-CT synthesis for improved synthetic CT generation: an iterative refinement and knowledge distillation approach

Objective.Deep learning has shown promise in generating synthetic CT (sCT) from magnetic resonance imaging (MRI). However, the misalignment between MRIs and CTs has not been adequately addressed, leading to reduced prediction accuracy and potential harm to patients due to the generative adversarial network (GAN)hallucination phenomenon. This work proposes a novel approach to mitigate misalignment and improve sCT generation.Approach.Our approach has two stages: iterative refinement and knowledge distillation. First, we iteratively refine registration and synthesis by leveraging their complementary nature. In each iteration, we register CT to the sCT from the previous iteration, generating a more aligned deformed CT (dCT). We train a new model on the refined 〈dCT, MRI〉 pairs to enhance synthesis. Second, we distill knowledge by creating a target CT (tCT) that combines sCT and dCT images from the previous iterations. This further improves alignment beyond the individual sCT and dCT images. We train a new model with the 〈tCT, MRI〉 pairs to transfer insights from multiple models into this final knowledgeable model.Main results.Our method outperformed conditional GANs on 48 head and neck cancer patients. It reduced hallucinations and improved accuracy in geometry (3% ↑ Dice), intensity (16.7% ↓ MAE), and dosimetry (1% ↑γ3%3mm). It also achieved <1% relative dose difference for specific dose volume histogram points.Significance.This pioneering approach for addressing misalignment shows promising performance in MRI-to-CT synthesis for MRI-only planning. It could be applied to other modalities like cone beam computed tomography and tasks such as organ contouring.

Bony structure enhanced synthetic CT generation using Dixon sequences for pelvis MR-only radiotherapy

BACKGROUND

MRI-only radiotherapy planning (MROP) is beneficial to patients by avoiding MRI/CT registration errors, simplifying the radiation treatment simulation workflow and reducing exposure to ionizing radiation. MRI is the primary imaging modality for soft tissue delineation. Treatment planning CTs (i.e., CT simulation scan) are redundant if a synthetic CT (sCT) can be generated from the MRI to provide the patient positioning and electron density information. Unsupervised deep learning (DL) models like CycleGAN are widely used in MR-to-sCT conversion, when paired patient CT and MR image datasets are not available for model training. However, compared to supervised DL models, they cannot guarantee anatomic consistency, especially around bone.

PURPOSE

The purpose of this work was to improve the sCT accuracy generated from MRI around bone for MROP.

METHODS

To generate more reliable bony structures on sCT images, we proposed to add bony structure constraints in the unsupervised CycleGAN model's loss function and leverage Dixon constructed fat and in-phase (IP) MR images. Dixon images provide better bone contrast than T2-weighted images as inputs to a modified multi-channel CycleGAN. A private dataset with a total of 31 prostate cancer patients were used for training (20) and testing (11).

RESULTS

We compared model performance with and without bony structure constraints using single- and multi-channel inputs. Among all the models, multi-channel CycleGAN with bony structure constraints had the lowest mean absolute error, both inside the bone and whole body (50.7 and 145.2 HU). This approach also resulted in the highest Dice similarity coefficient (0.88) of all bony structures compared with the planning CT.

CONCLUSION

Modified multi-channel CycleGAN with bony structure constraints, taking Dixon-constructed fat and IP images as inputs, can generate clinically suitable sCT images in both bone and soft tissue. The generated sCT images have the potential to be used for accurate dose calculation and patient positioning in MROP radiation therapy.

Technical note: Evaluation of deep learning based synthetic CTs clinical readiness for dose and NTCP driven head and neck adaptive proton therapy

BACKGROUND

Adaptive proton therapy workflows rely on accurate imaging throughout the treatment course. Our centre currently utilizes weekly repeat CTs (rCTs) for treatment monitoring and plan adaptations. However, deep learning-based methods have recently shown to successfully correct CBCT images, which suffer from severe imaging artifacts, and generate high quality synthetic CT (sCT) images which enable CBCT-based proton dose calculations.

PURPOSE

To compare daily CBCT-based sCT images to planning CTs (pCT) and rCTs of head and neck (HN) cancer patients to investigate the dosimetric accuracy of CBCT-based sCTs in a scenario mimicking actual clinical practice.

METHODS

Data of 56 HN cancer patients, previously treated with proton therapy was used to generate 1.962 sCT images, using a previously developed and trained deep convolutional neural network. Clinical IMPT treatment plans were recalculated on the pCT, weekly rCTs and daily sCTs. The dosimetric accuracy of sCTs was compared to same day rCTs and the initial planning CT. As a reference, rCTs were also compared to pCTs. The dose difference between sCTs and rCTs/pCT was quantified by calculating the D98 difference for target volumes and Dmean difference for organs-at-risk. To investigate the clinical relevancy of possible dose differences, NTCP values were calculated for dysphagia and xerostomia.

RESULTS

For target volumes, only minor dose differences were found for sCT versus rCT and sCT versus pCT, with dose differences mostly within ±1.5%. Larger dose differences were observed in OARs, where a general shift towards positive differences was found, with the largest difference in the left parotid gland. Delta NTCP values for grade 2 dysphagia and xerostomia were within ±2.5% for 90% of the sCTs.

CONCLUSIONS

Target doses showed high similarity between rCTs and sCTs. Further investigations are required to identify the origin of the dose differences at OAR levels and its relevance in clinical decision making.

Artificial general intelligence for radiation oncology

The emergence of artificial general intelligence (AGI) is transforming radiation oncology. As prominent vanguards of AGI, large language models (LLMs) such as GPT-4 and PaLM 2 can process extensive texts and large vision models (LVMs) such as the Segment Anything Model (SAM) can process extensive imaging data to enhance the efficiency and precision of radiation therapy. This paper explores full-spectrum applications of AGI across radiation oncology including initial consultation, simulation, treatment planning, treatment delivery, treatment verification, and patient follow-up. The fusion of vision data with LLMs also creates powerful multimodal models that elucidate nuanced clinical patterns. Together, AGI promises to catalyze a shift towards data-driven, personalized radiation therapy. However, these models should complement human expertise and care. This paper provides an overview of how AGI can transform radiation oncology to elevate the standard of patient care in radiation oncology, with the key insight being AGI's ability to exploit multimodal clinical data at scale.

Proton range uncertainty caused by synthetic computed tomography generated with deep learning from pelvic magnetic resonance imaging

BACKGROUND

In proton therapy, it is disputed whether synthetic computed tomography (sCT), derived from magnetic resonance imaging (MRI), permits accurate dose calculations. On the one hand, an MRI-only workflow could eliminate errors caused by, e.g., MRI-CT registration. On the other hand, the extra error would be induced due to an sCT generation model. This work investigated the systematic and random model error induced by sCT generation of a widely discussed deep learning model, pix2pix.

MATERIAL AND METHODS

An open-source image dataset of 19 patients with cancer in the pelvis was employed and split into 10, 5, and 4 for training, testing, and validation of the model, respectively. Proton pencil beams (200 MeV) were simulated on the real CT and generated sCT using the tool for particle simulation (TOPAS). Monte Carlo (MC) dropout was used for error estimation (50 random sCT samples). Systematic and random model errors were investigated for sCT generation and dose calculation on sCT.

RESULTS

For sCT generation, random model error near the edge of the body (∼200 HU) was higher than that within the body (∼100 HU near the bone edge and <10 HU in soft tissue). The mean absolute error (MAE) was 49 ± 5, 191 ± 23, and 503 ± 70 HU for the whole body, bone, and air in the patient, respectively. Random model errors of the proton range were small (<0.2 mm) for all spots and evenly distributed throughout the proton fields. Systematic errors of the proton range were -1.0(±2.2) mm and 0.4(±0.9)%, respectively, and were unevenly distributed within the proton fields. For 4.5% of the spots, large errors (>5 mm) were found, which may relate to MRI-CT mismatch due to, e.g., registration, MRI distortion anatomical changes, etc.

CONCLUSION

The sCT model was shown to be robust, i.e., had a low random model error. However, further investigation to reduce and even predict and manage systematic error is still needed for future MRI-only proton therapy.

The Applications and Pitfalls of Cone-Beam Computed Tomography-Based Synthetic Computed Tomography for Adaptive Evaluation in Pencil-Beam Scanning Proton Therapy

PURPOSE

The study evaluates the efficacy of cone-beam computed tomography (CBCT)-based synthetic CTs (sCT) as a potential alternative to verification CT (vCT) for enhanced treatment monitoring and early adaptation in proton therapy.

METHODS

Seven common treatment sites were studied. Two sets of sCT per case were generated: direct-deformed (DD) sCT and image-correction (IC) sCT. The image qualities and dosimetric impact of the sCT were compared to the same-day vCT.

RESULTS

The sCT agreed with vCT in regions of homogeneous tissues such as the brain and breast; however, notable discrepancies were observed in the thorax and abdomen. The sCT outliers existed for DD sCT when there was an anatomy change and for IC sCT in low-density regions. The target coverage exhibited less than a 5% variance in most DD and IC sCT cases when compared to vCT. The Dmax of serial organ-at-risk (OAR) in sCT plans shows greater deviation from vCT than small-volume dose metrics (D0.1cc). The parallel OAR volumetric and mean doses remained consistent, with average deviations below 1.5%.

CONCLUSION

The use of sCT enables precise treatment and prompt early adaptation for proton therapy. The quality assurance of sCT is mandatory in the early stage of clinical implementation.

Synthetic CTs for MRI-only brain RT treatment: integration of immobilization systems

PURPOSE

Auxiliary devices such as immobilization systems should be considered in synthetic CT (sCT)-based treatment planning (TP) for MRI-only brain radiotherapy (RT). A method for auxiliary device definition in the sCT is introduced, and its dosimetric impact on the sCT-based TP is addressed.

METHODS

T1-VIBE DIXON was acquired in an RT setup. Ten datasets were retrospectively used for sCT generation. Silicone markers were used to determine the auxiliary devices' relative position. An auxiliary structure template (AST) was created in the TP system and placed manually on the MRI. Various RT mask characteristics were simulated in the sCT and investigated by recalculating the CT-based clinical plan on the sCT. The influence of auxiliary devices was investigated by creating static fields aimed at artificial planning target volumes (PTVs) in the CT and recalculated in the sCT. The dose covering 50% of the PTV (D50) deviation percentage between CT-based/recalculated plan (∆D50[%]) was evaluated.

RESULTS

Defining an optimal RT mask yielded a ∆D50[%] of 0.2 ± 1.03% for the PTV and between -1.6 ± 3.4% and 1.1 ± 2.0% for OARs. Evaluating each static field, the largest ∆D50[%] was delivered by AST positioning inaccuracy (max: 3.5 ± 2.4%), followed by the RT table (max: 3.6 ± 1.2%) and the RT mask (max: 3.0 ± 0.8% [anterior], 1.6 ± 0.4% [rest]). No correlation between ∆D50[%] and beam depth was found for the sum of opposing beams, except for (45° + 315°).

CONCLUSION

This study evaluated the integration of auxiliary devices and their dosimetric influence on sCT-based TP. The AST can be easily integrated into the sCT-based TP. Further, we found that the dosimetric impact was within an acceptable range for an MRI-only workflow.

Ensemble learning and personalized training for the improvement of unsupervised deep learning-based synthetic CT reconstruction

BACKGROUND

The growing adoption of magnetic resonance imaging (MRI)-guided radiation therapy (RT) platforms and a focus on MRI-only RT workflows have brought the technical challenge of synthetic computed tomography (sCT) reconstruction to the forefront. Unpaired-data deep learning-based approaches to the problem offer the attractive characteristic of not requiring paired training data, but the gap between paired- and unpaired-data results can be limiting.

PURPOSE

We present two distinct approaches aimed at improving unpaired-data sCT reconstruction results: a cascade ensemble that combines multiple models and a personalized training strategy originally designed for the paired-data setting.

METHODS

Comparisons are made between the following models: (1) the paired-data fully convolutional DenseNet (FCDN), (2) the FCDN with the Intentional Deep Overfit Learning (IDOL) personalized training strategy, (3) the unpaired-data CycleGAN, (4) the CycleGAN with the IDOL training strategy, and (5) the CycleGAN as an intermediate model in a cascade ensemble approach. Evaluation of the various models over 25 total patients is carried out using a five-fold cross-validation scheme, with the patient-specific IDOL models being trained for the five patients of fold 3, chosen at random.

RESULTS

In both the paired- and unpaired-data settings, adopting the IDOL training strategy led to improvements in the mean absolute error (MAE) between true CT images and sCT outputs within the body contour (mean improvement, paired- and unpaired-data approaches, respectively: 38%, 9%) and in regions of bone (52%, 5%), the peak signal-to-noise ratio (PSNR; 15%, 7%), and the structural similarity index (SSIM; 6%, <1%). The ensemble approach offered additional benefits over the IDOL approach in all three metrics (mean improvement over unpaired-data approach in fold 3; MAE: 20%; bone MAE: 16%; PSNR: 10%; SSIM: 2%), and differences in body MAE between the ensemble approach and the paired-data approach are statistically insignificant.

CONCLUSIONS

We have demonstrated that both a cascade ensemble approach and a personalized training strategy designed initially for the paired-data setting offer significant improvements in image quality metrics for the unpaired-data sCT reconstruction task. Closing the gap between paired- and unpaired-data approaches is a step toward fully enabling these powerful and attractive unpaired-data frameworks.

Synthetic CT in Carbon Ion Radiotherapy of the Abdominal Site

The generation of synthetic CT for carbon ion radiotherapy (CIRT) applications is challenging, since high accuracy is required in treatment planning and delivery, especially in an anatomical site as complex as the abdomen. Thirty-nine abdominal MRI-CT volume pairs were collected and a three-channel cGAN (accounting for air, bones, soft tissues) was used to generate sCTs. The network was tested on five held-out MRI volumes for two scenarios: (i) a CT-based segmentation of the MRI channels, to assess the quality of sCTs and (ii) an MRI manual segmentation, to simulate an MRI-only treatment scenario. The sCTs were evaluated by means of similarity metrics (e.g., mean absolute error, MAE) and geometrical criteria (e.g., dice coefficient). Recalculated CIRT plans were evaluated through dose volume histogram, gamma analysis and range shift analysis. The CT-based test set presented optimal MAE on bones (86.03 ± 10.76 HU), soft tissues (55.39 ± 3.41 HU) and air (54.42 ± 11.48 HU). Higher values were obtained from the MRI-only test set (MAE_BONE = 154.87 ± 22.90 HU). The global gamma pass rate reached 94.88 ± 4.9% with 3%/3 mm, while the range shift reached a median (IQR) of 0.98 (3.64) mm. The three-channel cGAN can generate acceptable abdominal sCTs and allow for CIRT dose recalculations comparable to the clinical plans.

Streaking artifact reduction for CBCT-based synthetic CT generation in adaptive radiotherapy

BACKGROUND

Cone-beam computed tomography (CBCT) is widely used for daily image guidance in radiation therapy, enhancing the reproducibility of patient setup. However, its application in adaptive radiotherapy (ART) is limited by many imaging artifacts and inaccurate Hounsfield units (HUs). The correction of CBCT image is necessary and of great value for CBCT-based ART.

PURPOSE

To explore the synthetic CT (sCT) generation from CBCT images of thorax and abdomen patients, which usually surfer from serious artifacts duo to organ state changes. In this study, a streaking artifact reduction network (SARN) is proposed to reduce artifacts and combine with cycleGAN to generate high-quality sCT images from CBCT and achieve an accurate dose calculation.

METHODS

The proposed SARN was trained in a self-supervised manner. Artifact-CT images were generated from planning CT by random deformation and projection replacement, and SARN was trained based on paired artifact-CT and CT images. The planning CT and CBCT images of 260 patients with cancer, including 120 thoracic and 140 abdominal CT scans, were used to train and evaluate neural networks. The CBCT images of another 12 patients in late treatment fractions, which contained large anatomy changes, were also tested by trained models. The trained models include commonly used U-Net, cycleGAN, attention-gated cycleGAN (cycAT), and cascade models combined SARN with cycleGAN or cycAT. The generated sCT images were compared in terms of image quality and dose calculation accuracy.

RESULTS

The sCT images generated by SARN combined with cycleGAN and cycAT showed the best image quality, removed the most artifacts, and retained the normal anatomical structure. The SARN+cycleGAN performed best in streaking artifacts removal with the maximum percent integrity uniformity (PIU_m ) of 91.0% and minimum standard deviation (SD) of 35.4 HU for delineated artifact regions among all models. The mean absolute error (MAE) of CBCT images in the thorax and abdomen were 71.6 and 55.2 HU, respectively, using planning CT images after deformable registration as ground truth. Compared with CBCT, the thoracic and abdominal sCT images generated by each model had significantly improved image quality with smaller MAE (p < 0.05). The SARN+cycAT obtained the minimum MAEs of 42.5 HU in the thorax while SARN+cycleGAN got the minimum MAEs of 32.0 HU in the abdomen. The sCT generated by U-Net had a remarkably lower anatomical structure accuracy compared with the other models. The thoracic and abdominal sCT images generated by SARN+cycleGAN showed optimal dose calculation accuracy with gamma passing rates (2 mm/2%) of 98.2% and 96.9%, respectively.

CONCLUSIONS

The proposed SARN can reduce serious streaking artifacts in CBCT images. The SARN combined with cycleGAN can generate high-quality sCT images with fewer artifacts, high-accuracy HU values, and accurate anatomical structures, thus providing reliable dose calculation in ART.

A deep learning approach to generate synthetic CT in low field MR-guided radiotherapy for lung cases

INTRODUCTION

This study aims to apply a conditional Generative Adversarial Network (cGAN) to generate synthetic Computed Tomography (sCT) from 0.35 Tesla Magnetic Resonance (MR) images of the thorax.

METHODS

Sixty patients treated for lung lesions were enrolled and divided into training (32), validation (8), internal (10,T_A) and external (10,T_B) test set. Image accuracy of generated sCT was evaluated computing the mean absolute (MAE) and mean error (ME) with respect the original CT. Three treatment plans were calculated for each patient considering MRI as reference image: original CT, sCT (pure sCT) and sCT with GTV density override (hybrid sCT) were used as Electron Density (ED) map. Dose accuracy was evaluated comparing treatment plans in terms of gamma analysis and Dose Volume Histogram (DVH) parameters.

RESULTS

No significant difference was observed between the test sets for image and dose accuracy parameters. Considering the whole test cohort, a MAE of 54.9 ± 10.5 HU and a ME of 4.4 ± 7.4 HU was obtained. Mean gamma passing rates for 2%/2mm, and 3%/3mm tolerance criteria were 95.5 ± 5.9% and 98.2 ± 4.1% for pure sCT, 96.1 ± 5.1% and 98.5 ± 3.9% for hybrid sCT: the difference between the two approaches was significant (p = 0.01). As regards DVH analysis, differences in target parameters estimation were found to be within 5% using hybrid approach and 20% using pure sCT.

CONCLUSION

The DL algorithm here presented can generate sCT images in the thorax with good image and dose accuracy, especially when the hybrid approach is used. The algorithm does not suffer from inter-scanner variability, making feasible the implementation of MR-only workflows for palliative treatments.

Deep learning-based 4D-synthetic CTs from sparse-view CBCTs for dose calculations in adaptive proton therapy

BACKGROUND

Time-resolved 4D cone beam-computed tomography (4D-CBCT) allows a daily assessment of patient anatomy and respiratory motion. However, 4D-CBCTs suffer from imaging artifacts that affect the CT number accuracy and prevent accurate proton dose calculations. Deep learning can be used to correct CT numbers and generate synthetic CTs (sCTs) that can enable CBCT-based proton dose calculations.

PURPOSE

In this work, sparse view 4D-CBCTs were converted into 4D-sCT utilizing a deep convolutional neural network (DCNN). 4D-sCTs were evaluated in terms of image quality and dosimetric accuracy to determine if accurate proton dose calculations for adaptive proton therapy workflows of lung cancer patients are feasible.

METHODS

A dataset of 45 thoracic cancer patients was utilized to train and evaluate a DCNN to generate 4D-sCTs, based on sparse view 4D-CBCTs reconstructed from projections acquired with a 3D acquisition protocol. Mean absolute error (MAE) and mean error were used as metrics to evaluate the image quality of single phases and average 4D-sCTs against 4D-CTs acquired on the same day. The dosimetric accuracy was checked globally (gamma analysis) and locally for target volumes and organs-at-risk (OARs) (lung, heart, and esophagus). Furthermore, 4D-sCTs were also compared to 3D-sCTs. To evaluate CT number accuracy, proton radiography simulations in 4D-sCT and 4D-CTs were compared in terms of range errors. The clinical suitability of 4D-sCTs was demonstrated by performing a 4D dose reconstruction using patient specific treatment delivery log files and breathing signals.

RESULTS

4D-sCTs resulted in average MAEs of 48.1 ± 6.5 HU (single phase) and 37.7 ± 6.2 HU (average). The global dosimetric evaluation showed gamma pass ratios of 92.3% ± 3.2% (single phase) and 94.4% ± 2.1% (average). The clinical target volume showed high agreement in D₉₈ between 4D-CT and 4D-sCT, with differences below 2.4% for all patients. Larger dose differences were observed in mean doses of OARs (up to 8.4%). The comparison with 3D-sCTs showed no substantial image quality and dosimetric differences for the 4D-sCT average. Individual 4D-sCT phases showed slightly lower dosimetric accuracy. The range error evaluation revealed that lung tissues cause range errors about three times higher than the other tissues.

CONCLUSION

In this study, we have investigated the accuracy of deep learning-based 4D-sCTs for daily dose calculations in adaptive proton therapy. Despite image quality differences between 4D-sCTs and 3D-sCTs, comparable dosimetric accuracy was observed globally and locally. Further improvement of 3D and 4D lung sCTs could be achieved by increasing CT number accuracy in lung tissues.

Quality assurance for MRI-only radiation therapy: A voxel-wise population-based methodology for image and dose assessment of synthetic CT generation methods

The quality assurance of synthetic CT (sCT) is crucial for safe clinical transfer to an MRI-only radiotherapy planning workflow. The aim of this work is to propose a population-based process assessing local errors in the generation of sCTs and their impact on dose distribution. For the analysis to be anatomically meaningful, a customized interpatient registration method brought the population data to the same coordinate system. Then, the voxel-based process was applied on two sCT generation methods: a bulk-density method and a generative adversarial network. The CT and MRI pairs of 39 patients treated by radiotherapy for prostate cancer were used for sCT generation, and 26 of them with delineated structures were selected for analysis. Voxel-wise errors in sCT compared to CT were assessed for image intensities and dose calculation, and a population-based statistical test was applied to identify the regions where discrepancies were significant. The cumulative histograms of the mean absolute dose error per volume of tissue were computed to give a quantitative indication of the error for each generation method. Accurate interpatient registration was achieved, with mean Dice scores higher than 0.91 for all organs. The proposed method produces three-dimensional maps that precisely show the location of the major discrepancies for both sCT generation methods, highlighting the heterogeneity of image and dose errors for sCT generation methods from MRI across the pelvic anatomy. Hence, this method provides additional information that will assist with both sCT development and quality control for MRI-based planning radiotherapy.

Classical and Learned MR to Pseudo-CT Mappings for Accurate Transcranial Ultrasound Simulation

Model-based treatment planning for transcranial ultrasound therapy typically involves mapping the acoustic properties of the skull from an X-ray computed tomography (CT) image of the head. Here, three methods for generating pseudo-CT (pCT) images from magnetic resonance (MR) images were compared as an alternative to CT. A convolutional neural network (U-Net) was trained on paired MR-CT images to generate pCT T images from either T1-weighted or zero-echo time (ZTE) MR images (denoted tCT and zCT, respectively). A direct mapping from ZTE to pCT was also implemented (denoted cCT). When comparing the pCT and ground-truth CT images for the test set, the mean absolute error was 133, 83, and 145 Hounsfield units (HU) across the whole head, and 398, 222, and 336 HU within the skull for the tCT, zCT, and cCT images, respectively. Ultrasound simulations were also performed using the generated pCT images and compared to simulations based on CT. An annular array transducer was used targeting the visual or motor cortex. The mean differences in the simulated focal pressure, focal position, and focal volume were 9.9%, 1.5 mm, and 15.1% for simulations based on the tCT images; 5.7%, 0.6 mm, and 5.7% for the zCT; and 6.7%, 0.9 mm, and 12.1% for the cCT. The improved results for images mapped from ZTE highlight the advantage of using imaging sequences, which improves the contrast of the skull bone. Overall, these results demonstrate that acoustic simulations based on MR images can give comparable accuracy to those based on CT.

Development of an unsupervised cycle contrastive unpaired translation network for MRI-to-CT synthesis

Purpose

The purpose of this work is to develop and evaluate a novel cycle‐contrastive unpaired translation network (cycleCUT) for synthetic computed tomography (sCT) generation from T1‐weighted magnetic resonance images (MRI).

Methods

The cycleCUT proposed in this work integrated the contrastive learning module from contrastive unpaired translation network (CUT) into the cycle‐consistent generative adversarial network (cycleGAN) framework to effectively achieve unsupervised CT synthesis from MRI. The diagnostic MRI and radiotherapy planning CT images of 24 brain cancer patients were obtained and reshuffled to train the network. For comparison, the traditional cycleGAN and CUT were also implemented. The sCT images were then imported into a treatment planning system to verify their feasibility for radiotherapy planning. The mean absolute error (MAE), peak signal‐to‐noise ratio (PSNR), and structural similarity index (SSIM) between the sCT and the corresponding real CT images were calculated. Gamma analysis between sCT‐ and CT‐based dose distributions was also conducted.

Results

Quantitative evaluation of an independent test set of six patients showed that the average MAE was 69.62 ± 5.68 Hounsfield Units (HU) for the proposed cycleCUT, significantly (p‐value < 0.05) lower than that for cycleGAN (77.02 ± 6.00 HU) and CUT (78.05 ± 8.29). The average PSNR was 28.73 ± 0.46 decibels (dB) for cycleCUT, significantly larger than that for cycleGAN (27.96 ± 0.49 dB) and CUT (27.95 ± 0.69 dB). The average SSIM for cycleCUT (0.918 ± 0.012) was also significantly higher than that for cycleGAN (0.906 ± 0.012) and CUT (0.903 ± 0.015). Regarding gamma analysis, cycleCUT achieved the highest passing rate (97.95 ± 1.24% at the 2%/2 mm criteria and 10% dose threshold) but was not significantly different from the others.

Conclusion

The proposed cycleCUT could be effectively trained using unaligned image data, and could generate better sCT images than cycleGAN and CUT in terms of HU number accuracy and fine structural details.

Emergence of MR-Linac in Radiation Oncology: Successes and Challenges of Riding on the MRgRT Bandwagon

The special issue of JCM on “Advances of MRI in Radiation Oncology” provides a unique forum for scientific literature related to MR imaging in radiation oncology. This issue covered many aspects, such as MR technology, motion management, economics, soft-tissue–air interface issues, and disease sites such as the pancreas, spine, sarcoma, prostate, head and neck, and rectum from both camps—the Unity and MRIdian systems. This paper provides additional information on the success and challenges of the two systems. A challenging aspect of this technology is low throughput and the monumental task of education and training that hinders its use for the majority of therapy centers. Additionally, the cost of this technology is too high for most institutions, and hence widespread use is still limited. This article highlights some of the difficulties and how to resolve them.

Adaptive Proton Therapy of Pediatric Head and Neck Cases Using MRI-Based Synthetic CTs: Initial Experience of the Prospective KiAPT Study

BACKGROUND AND PURPOSE

Interfractional anatomical changes might affect the outcome of proton therapy (PT). We aimed to prospectively evaluate the role of Magnetic Resonance Imaging (MRI) based adaptive PT for children with tumors of the head and neck and base of skull.

METHODS

MRI verification images were acquired at half of the treatment course. A synthetic computed tomography (CT) image was created using this MRI and a deformable image registration (DIR) to the reference MRI. The methodology was verified with in-silico phantoms and validated using a clinical case with a shrinking cystic hygroma on the basis of dosimetric quantities of contoured structures. The dose distributions on the verification X-ray CT and on the synthetic CT were compared with a gamma-index test using global 2 mm/2% criteria.

RESULTS

Regarding the clinical validation case, the gamma-index pass rate was 98.3%. Eleven patients were included in the clinical study. The most common diagnosis was rhabdomyosarcoma (73%). Craniofacial tumor site was predominant in 64% of patients, followed by base of skull (18%). For one individual case the synthetic CT showed an increase in the median D2 and Dmax dose on the spinal cord from 20.5 GyRBE to 24.8 GyRBE and 14.7 GyRBE to 25.1 GyRBE, respectively. Otherwise, doses received by OARs remained relatively stable. Similarly, the target volume coverage seen by D95% and V95% remained unchanged.

CONCLUSIONS

The method of transferring anatomical changes from MRIs to a synthetic CTs was successfully implemented and validated with simple, commonly available tools. In the frame of our early results on a small cohort, no clinical relevant deterioration for neither PTV coverage nor an increased dose burden to OARs occurred. However, the study will be continued to identify a pediatric patient cohort, which benefits from adaptive treatment planning.

Evaluating Proton Dose and Associated Range Uncertainty Using Daily Cone-Beam CT

Purpose

This study aimed to quantitatively evaluate the range uncertainties that arise from daily cone-beam CT (CBCT) images for proton dose calculation compared to CT using a measurement-based technique.

Methods

For head and thorax phantoms, wedge-shaped intensity-modulated proton therapy (IMPT) treatment plans were created such that the gradient of the wedge intersected and was measured with a 2D ion chamber array. The measured 2D dose distributions were compared with 2D dose planes extracted from the dose distributions using the IMPT plan calculated on CT and CBCT. Treatment plans of a thymoma cancer patient treated with breath-hold (BH) IMPT were recalculated on 28 CBCTs and 9 CTs, and the resulting dose distributions were compared.

Results

The range uncertainties for the head phantom were determined to be 1.2% with CBCT, compared to 0.5% for CT, whereas the range uncertainties for the thorax phantom were 2.1% with CBCT, compared to 0.8% for CT. The doses calculated on CBCT and CT were similar with similar anatomy changes. For the thymoma patient, the primary source of anatomy change was the BH uncertainty, which could be up to 8 mm in the superior-inferior (SI) direction.

Conclusion

We developed a measurement-based range uncertainty evaluation method with high sensitivity and used it to validate the accuracy of CBCT-based range and dose calculation. Our study demonstrated that the CBCT-based dose calculation could be used for daily dose validation in selected proton patients.

Unsupervised pseudo CT generation using heterogenous multicentric CT/MR images and CycleGAN: Dosimetric assessment for 3D conformal radiotherapy

PURPOSE

Absorbed dose calculation in magnetic resonance-guided radiation therapy (MRgRT) is commonly based on pseudo CT (pCT) images. This study investigated the feasibility of unsupervised pCT generation from MRI using a cycle generative adversarial network (CycleGAN) and a heterogenous multicentric dataset. A dosimetric analysis in three-dimensional conformal radiotherapy (3DCRT) planning was also performed.

MATERIAL AND METHODS

Overall, 87 T1-weighted and 102 T2-weighted MR images alongside with their corresponding computed tomography (CT) images of brain cancer patients from multiple centers were used. Initially, images underwent a number of preprocessing steps, including rigid registration, novel CT Masker, N4 bias field correction, resampling, resizing, and rescaling. To overcome the gradient vanishing problem, residual blocks and mean squared error (MSE) loss function were utilized in the generator and in both networks (generator and discriminator), respectively. The CycleGAN was trained and validated using 70 T1 and 80 T2 randomly selected patients in an unsupervised manner. The remaining patients were used as a holdout test set to report final evaluation metrics. The generated pCTs were validated in the context of 3DCRT.

RESULTS

The CycleGAN model using masked T2 images achieved better performance with a mean absolute error (MAE) of 61.87 ± 22.58 HU, peak signal to noise ratio (PSNR) of 27.05 ± 2.25 (dB), and structural similarity index metric (SSIM) of 0.84 ± 0.05 on the test dataset. T1-weighted MR images used for dosimetric assessment revealed a gamma index of 3%, 3 mm, 2%, 2 mm and 1%, 1 mm with acceptance criteria of 98.96% ± 1.1%, 95% ± 3.68%, 90.1% ± 6.05%, respectively. The DVH differences between CTs and pCTs were within 2%.

CONCLUSIONS

A promising pCT generation model capable of handling heterogenous multicenteric datasets was proposed. All MR sequences performed competitively with no significant difference in pCT generation. The proposed CT Masker proved promising in improving the model accuracy and robustness. There was no significant difference between using T1-weighted and T2-weighted MR images for pCT generation.

Toward MR-only proton therapy planning for pediatric brain tumors: synthesis of relative proton stopping power images with multiple sequence MRI and development of an online quality assurance tool

PURPOSE

To generate synthetic relative proton-stopping-power (sRPSP) images from MRI sequence(s) and develop an online quality assurance (QA) tool for sRPSP to facilitate safe integration of MR-only proton planning into clinical practice.

MATERIALS AND METHODS

Planning CT and MR images of 195 pediatric brain tumor patients were utilized (training: 150, testing: 45). Seventeen consistent-cycle Generative Adversarial Network (ccGAN) models were trained separately using paired CT-converted RPSP and MRI datasets to transform a subject's MRI into sRPSP. T1-weighted (T1W), T2-weighted (T2W), and FLAIR MRI were permutated to form 17 combinations, with or without preprocessing, for determining the optimal training sequence(s). For evaluation, sRPSP images were converted to synthetic CT (sCT) and compared to the real CT in terms of mean absolute error (MAE) in HU. For QA, sCT was deformed and compared to a reference template built from training dataset to produce a flag map, highlighting pixels that deviate by >100 HU and fall outside the mean ± standard deviation reference intensity. The gamma intensity analysis (10%/3mm) of the deformed sCT against the QA template on the intensity difference was investigated as a surrogate of sCT accuracy.

RESULTS

The sRPSP images generated from a single T1W or T2W sequence outperformed that generated from multi-MRI sequences in terms of MAE (all P<0.05). Preprocessing with N4 bias and histogram matching reduced MAE of T2W MRI-based sCT (54±21 HU vs. 42±13 HU, P = .002). The gamma intensity analysis of sCT against the QA template was highly correlated with the MAE of sCT against the real CT in the testing cohort (r = -0.89 for T1W sCT; r = -0.93 for T2W sCT).

CONCLUSION

Accurate sRPSP images can be generated from T1W/T2W MRI for proton planning. A QA tool highlights regions of inaccuracy, flagging problematic cases unsuitable for clinical use. This article is protected by copyright. All rights reserved.

An MRI sequence independent convolutional neural network for synthetic head CT generation in proton therapy

A magnetic resonance imaging (MRI) sequence independent deep learning technique was developed and validated to generate synthetic computed tomography (sCT) scans for MR guided proton therapy. 47 meningioma patients previously undergoing proton therapy based on pencil beam scanning were divided into training (33), validation (6), and test (8) cohorts. T₁, T₂, and contrast enhanced T₁ (T1CM) MRI sequences were used in combination with the planning CT (pCT) data to train a 3D U-Net architecture with ResNet-Blocks. A hyperparameter search was performed including two loss functions, two group sizes of normalisation, and depth of the network. Training outcome was compared between models trained for each individual MRI sequence and for all sequences combined. The performance was evaluated based on a metric and dosimetric analysis as well as spot difference maps. Furthermore, the influence of immobilisation masks that are not visible on MRIs was investigated. Based on the hyperparameter search, the final model was trained with fixed features per group for the group normalisation, six down-convolution steps, an input size of 128×192×192, and feature loss. For the test dataset for body/bone the mean absolute error (MAE) values were on average 79.8/216.3Houndsfield unit (HU) when trained using T1 images, 71.1/186.1HU for T2, and 82.9/236.4HU for T1CM. The structural similarity metric (SSIM) ranged from 0.95 to 0.98 for all sequences. The investigated dose parameters of the target structures agreed within 1% between original proton treatment plans and plans recalculated on sCTs. The spot difference maps had peaks at ±0.2cm and for 98% of all spots the difference was less than 1cm. A novel MRI sequence independent sCT generator was developed, which suggests that the training phase of neural networks can be disengaged from specific MRI acquisition protocols. In contrast to previous studies, the patient cohort consisted exclusively of actual proton therapy patients (i.e. "real-world data").

Clinical suitability of deep learning based synthetic CTs for adaptive proton therapy of lung cancer

Purpose

Adaptive proton therapy (APT) of lung cancer patients requires frequent volumetric imaging of diagnostic quality. Cone‐beam CT (CBCT) can provide these daily images, but x‐ray scattering limits CBCT‐image quality and hampers dose calculation accuracy. The purpose of this study was to generate CBCT‐based synthetic CTs using a deep convolutional neural network (DCNN) and investigate image quality and clinical suitability for proton dose calculations in lung cancer patients.

Methods

A dataset of 33 thoracic cancer patients, containing CBCTs, same‐day repeat CTs (rCT), planning‐CTs (pCTs), and clinical proton treatment plans, was used to train and evaluate a DCNN with and without a pCT‐based correction method. Mean absolute error (MAE), mean error (ME), peak signal‐to‐noise ratio, and structural similarity were used to quantify image quality. The evaluation of clinical suitability was based on recalculation of clinical proton treatment plans. Gamma pass ratios, mean dose to target volumes and organs at risk, and normal tissue complication probabilities (NTCP) were calculated. Furthermore, proton radiography simulations were performed to assess the HU‐accuracy of sCTs in terms of range errors.

Results

On average, sCTs without correction resulted in a MAE of 34 ± 6 HU and ME of 4 ± 8 HU. The correction reduced the MAE to 31 ± 4HU (ME to 2 ± 4HU). Average 3%/3 mm gamma pass ratios increased from 93.7% to 96.8%, when the correction was applied. The patient specific correction reduced mean proton range errors from 1.5 to 1.1 mm. Relative mean target dose differences between sCTs and rCT were below ± 0.5% for all patients and both synthetic CTs (with/without correction). NTCP values showed high agreement between sCTs and rCT (<2%).

Conclusion

CBCT‐based sCTs can enable accurate proton dose calculations for APT of lung cancer patients. The patient specific correction method increased the image quality and dosimetric accuracy but had only a limited influence on clinically relevant parameters.

Deep learning methods to generate synthetic CT from MRI in radiotherapy: A literature review

PURPOSE

In radiotherapy, MRI is used for target volume and organs-at-risk delineation for its superior soft-tissue contrast as compared to CT imaging. However, MRI does not provide the electron density of tissue necessary for dose calculation. Several methods of synthetic-CT (sCT) generation from MRI data have been developed for radiotherapy dose calculation. This work reviewed deep learning (DL) sCT generation methods and their associated image and dose evaluation, in the context of MRI-based dose calculation.

METHODS

We searched the PubMed and ScienceDirect electronic databases from January 2010 to March 2021. For each paper, several items were screened and compiled in figures and tables.

RESULTS

This review included 57 studies. The DL methods were either generator-only based (45% of the reviewed studies), or generative adversarial network (GAN) architecture and its variants (55% of the reviewed studies). The brain and pelvis were the most commonly investigated anatomical localizations (39% and 28% of the reviewed studies, respectively), and more rarely, the head-and-neck (H&N) (15%), abdomen (10%), liver (5%) or breast (3%). All the studies performed an image evaluation of sCTs with a diversity of metrics, with only 36 studies performing dosimetric evaluations of sCT.

CONCLUSIONS

The median mean absolute errors were around 76 HU for the brain and H&N sCTs and 40 HU for the pelvis sCTs. For the brain, the mean dose difference between the sCT and the reference CT was <2%. For the H&N and pelvis, the mean dose difference was below 1% in most of the studies. Recent GAN architectures have advantages compared to generator-only, but no superiority was found in term of image or dose sCT uncertainties. Key challenges of DL-based sCT generation methods from MRI in radiotherapy is the management of movement for abdominal and thoracic localizations, the standardization of sCT evaluation, and the investigation of multicenter impacts.

Deep learning based synthetic-CT generation in radiotherapy and PET: A review

Recently,deep learning (DL)-based methods for the generation of synthetic computed tomography (sCT) have received significant research attention as an alternative to classical ones. We present here a systematic review of these methods by grouping them into three categories, according to their clinical applications: (i) to replace computed tomography in magnetic resonance (MR) based treatment planning, (ii) facilitate cone-beam computed tomography based image-guided adaptive radiotherapy, and (iii) derive attenuation maps for the correction of positron emission tomography. Appropriate database searching was performed on journal articles published between January 2014 and December 2020. The DL methods' key characteristics were extracted from each eligible study, and a comprehensive comparison among network architectures and metrics was reported. A detailed review of each category was given, highlighting essential contributions, identifying specific challenges, and summarizing the achievements. Lastly, the statistics of all the cited works from various aspects were analyzed, revealing the popularity and future trends and the potential of DL-based sCT generation. The current status of DL-based sCT generation was evaluated, assessing the clinical readiness of the presented methods.

Range probing as a quality control tool for CBCT-based synthetic CTs: In vivo application for head and neck cancer patients

Purpose

Cone‐beam CT (CBCT)‐based synthetic CTs (sCT) produced with a deep convolutional neural network (DCNN) show high image quality, suggesting their potential usability in adaptive proton therapy workflows. However, the nature of such workflows involving DCNNs prevents the user from having direct control over their output. Therefore, quality control (QC) tools that monitor the sCTs and detect failures or outliers in the generated images are needed.

This work evaluates the potential of using a range‐probing (RP)‐based QC tool to verify sCTs generated by a DCNN. Such a RP QC tool experimentally assesses the CT number accuracy in sCTs.

Methods

A RP QC dataset consisting of repeat CTs (rCT), CBCTs, and RP acquisitions of seven head and neck cancer patients was retrospectively assessed. CBCT‐based sCTs were generated using a DCNN. The CT number accuracy in the sCTs was evaluated by computing relative range errors between measured RP fields and RP field simulations based on rCT and sCT images.

Results

Mean relative range errors showed agreement between measured and simulated RP fields, ranging from −1.2% to 1.5% in rCTs, and from −0.7% to 2.7% in sCTs.

Conclusions

The agreement between measured and simulated RP fields suggests the suitability of sCTs for proton dose calculations. This outcome brings sCTs generated by DCNNs closer toward clinical implementation within adaptive proton therapy treatment workflows. The proposed RP QC tool allows for CT number accuracy assessment in sCTs and can provide means of in vivo range verification.

Cone beam CT based validation of neural network generated synthetic CTs for radiotherapy in the head region

PURPOSE

In the past years, many different neural network-based conversion techniques for synthesizing computed tomographys (sCTs) from MR images have been published. While the model's performance can be checked during the training against the test set, test datasets can never represent the whole population. Conversion errors can still occur for special cases, for example, for unusual anatomical situations. Therefore, the performance of sCT conversion needs to be verified on a patient specific level, especially in the absence of a planning CT (pCT). In this study, the capability of cone-beam CTs (CBCTs) for the validation of sCTs generated by a neural network was investigated.

METHODS

41 patients with tumors in the head region were selected. 20 of them were used for model training and 10 for validation. Different implementations of CycleGAN (with/without identity and feature loss) were used to generate sCTs. The pixel (MAE, RMSE, PSNR) and geometric error (DICE, Sensitivity, Specificity) values were reported to identify the best model. VMAT plans were created for the remaining 11 patients on the pCTs. These plans were re-calculated on sCTs and CBCTs. An automatic density overriding method ( C B C T RS ) and a population-based dose calculation method ( C B C T Pop ) were employed for CBCT-based dose calculation. The dose distributions were analysed using 3D global gamma analysis, applying a threshold of 10% with respect to the prescribed dose. Differences in DVH metrics for the PTV and the organs-at-risk were compared among the dose distributions based on pCTs, sCTs, and CBCTs.

RESULTS

The best model was the CycleGAN without identity and feature matching loss. Including the identity loss led to a metric decrease of 10% for DICE and a metric increase of 20-60 HU for MAE. Using the 2%/2 mm gamma criterion and pCT as reference, the mean gamma pass rates were 99.0  ±  0.4% for sCTs. Mean gamma pass rate values comparing pCT and CBCT were 99.0  ±  0.8% and 99.1  ±  0.8% for the C B C T RS and C B C T Pop , respectively. The mean gamma pass rates comparing sCT and CBCT resulted in 98.4  ±  1.6% and 99.2  ±  0.6% for C B C T RS and C B C T Pop , respectively. The differences between the gamma-pass-rates of the sCT and two CBCT-based methods were not significant. The majority of deviations of the investigated DVH metrices between sCTs and CBCTs were within 2%.

CONCLUSION

The dosimetric results demonstrate good agreement between sCT, CBCT, and pCT based calculations. A properly applied CBCT conversion method can serve as a tool for quality assurance procedures in an MR only radiotherapy workflow for head patients. Dosimetric deviations of DVH metrics between sCT and CBCTs of larger than 2% should be followed up. A systematic shift of approximately 1% should be taken into account when using the C B C T RS approach in an MR only workflow.

Facilitating MR-Guided Adaptive Proton Therapy in Children Using Deep Learning-Based Synthetic CT

Purpose

To determine whether self-attention cycle-generative adversarial networks (cycle-GANs), a novel deep-learning method, can generate accurate synthetic computed tomography (sCT) to facilitate adaptive proton therapy in children with brain tumors.

Materials and Methods

Both CT and T1-weighted magnetic resonance imaging (MRI) of 125 children (ages 1-20 years) with brain tumors were included in the training dataset. A model introducing a self-attention mechanism into the conventional cycle-GAN was created to enhance tissue interfaces and reduce noise. The test dataset consisted of 7 patients (ages 2-14 years) who underwent adaptive planning because of changes in anatomy discovered on MRI during proton therapy. The MRI during proton therapy-based sCT was compared with replanning CT (ground truth).

Results

The Hounsfield unit-mean absolute error was significantly reduced with self-attention cycle-GAN, as compared with conventional cycle-GAN (65.3 ± 13.9 versus 88.9 ± 19.3, P &lt; .01). The average 3-dimensional gamma passing rates (2%/2 mm criteria) for the original plan on the anatomy of the day and for the adapted plan were high (97.6% ± 1.2% and 98.9 ± 0.9%, respectively) when using sCT generated by self-attention cycle-GAN. The mean absolute differences in clinical target volume (CTV) receiving 95% of the prescription dose and 80% distal falloff along the beam axis were 1.1% ± 0.8% and 1.1 ± 0.9 mm, respectively. Areas of greatest dose difference were distal to the CTV and corresponded to shifts in distal falloff. Plan adaptation was appropriately triggered in all test patients when using sCT.

Conclusion

The novel cycle-GAN model with self-attention outperforms conventional cycle-GAN for children with brain tumors. Encouraging dosimetric results suggest that sCT generation can be used to identify patients who would benefit from adaptive replanning.

Training a deep neural network coping with diversities in abdominal and pelvic images of children and young adults for CBCT-based adaptive proton therapy

PURPOSE

To train a deep neural network for correcting abdominal and pelvic cone-beam computed tomography (CBCT) of children and young adults in the presence of diverse patient size, anatomic extent, and scan parameters.

MATERIALS AND METHODS

Pretreatment CBCT and planning/repeat CT image pairs from 64 children and young adults treated with proton therapy (aged 1-23 years) were analyzed. To evaluate the impact of anatomic extent in CBCT and data size in the training data, we compared the performance of three cycle-consistent generative adversarial network models that were separately trained by three datasets comprising abdominal (n = 21), pelvic (n = 29), and combined abdominal-pelvic image pairs (n = 50), respectively. The maximum body width of each patient was normalized to a fixed width before training and model application to reduce the impact of variations in body size. The corrected CBCT images by the three models were comparatively evaluated against the repeat CT closest in time to the CBCT (median gap, 0 days; range, 0-6 days) in HU accuracy, estimated dose distribution, and proton range.

RESULTS

The network model trained by the combined dataset significantly outperformed the abdomen and pelvis models in mean absolute HU error of the corrected CBCT from 14 testing patients (47 ± 7 HU versus 51 ± 8 HU; paired Wilcoxon signed-rank test, P < 0.01). The larger error (60 ± 7 HU) without the body-size normalization confirmed the efficacy of the preprocessing. The model trained with the combined dataset resulted in gamma passing rates of 98.5 ± 1.9% (2%/2 mm criterion) and the range (80% distal fall-off) differences from the reference within ±3 mm for 91.2 ± 11.5% beamlets.

CONCLUSION

Combining data from adjacent anatomic sites and normalizing age-dependent body sizes in children and young adults were beneficial in training a neural network to accurately estimate proton dose from CBCT despite limited training data size and anatomic diversities.

Effect of the Agglomerate Geometry on the Effective Electrical Conductivity of a Porous Electrode

The study of the microstructure of random heterogeneous materials, related to an electrochemical device, is relevant because their effective macroscopic properties, e.g., electrical or proton conductivity, are a function of their effective transport coefficients (ETC). The magnitude of ETC depends on the distribution and properties of the material phase. In this work, an algorithm is developed to generate stochastic two-phase (binary) image configurations with multiple geometries and polydispersed particle sizes. The recognizable geometry in the images is represented by the white phase dispersed and characterized by statistical descriptors (two-point and line-path correlation functions). Percolation is obtained for the geometries by identifying an infinite cluster to guarantee the connection between the edges of the microstructures. Finally, the finite volume method is used to determine the ETC. Agglomerate phase results show that the geometry with the highest local current distribution is the triangular geometry. In the matrix phase, the most significant results are obtained by circular geometry, while the lowest is obtained by the 3-sided polygon. The proposed methodology allows to establish criteria based on percolation and surface fraction to assure effective electrical conduction according to their geometric distribution; results provide an insight for the microstructure development with high projection to be used to improve the electrode of a Membrane Electrode Assembly (MEA).

Distant metastasis time to event analysis with CNNs in independent head and neck cancer cohorts

Deep learning models based on medical images play an increasingly important role for cancer outcome prediction. The standard approach involves usage of convolutional neural networks (CNNs) to automatically extract relevant features from the patient’s image and perform a binary classification of the occurrence of a given clinical endpoint. In this work, a 2D-CNN and a 3D-CNN for the binary classification of distant metastasis (DM) occurrence in head and neck cancer patients were extended to perform time-to-event analysis. The newly built CNNs incorporate censoring information and output DM-free probability curves as a function of time for every patient. In total, 1037 patients were used to build and assess the performance of the time-to-event model. Training and validation was based on 294 patients also used in a previous benchmark classification study while for testing 743 patients from three independent cohorts were used. The best network could reproduce the good results from 3-fold cross validation [Harrell’s concordance indices (HCIs) of 0.78, 0.74 and 0.80] in two out of three testing cohorts (HCIs of 0.88, 0.67 and 0.77). Additionally, the capability of the models for patient stratification into high and low-risk groups was investigated, the CNNs being able to significantly stratify all three testing cohorts. Results suggest that image-based deep learning models show good reliability for DM time-to-event analysis and could be used for treatment personalisation.

Synthetic dual-energy CT for MRI-only based proton therapy treatment planning using label-GAN

MRI-only treatment planning is highly desirable in the current proton radiation therapy workflow due to its appealing advantages such as bypassing MR-CT co-registration, avoiding x-ray CT exposure dose and reduced medical cost. However, MRI alone cannot provide stopping power ratio (SPR) information for dose calculations. Given that dual energy CT (DECT) can estimate SPR with higher accuracy than conventional single energy CT, we propose a deep learning-based method in this study to generate synthetic DECT (sDECT) from MRI to calculate SPR. Since the contrast difference between high-energy and low-energy CT (LECT) is important, and in order to accurately model this difference, we propose a novel label generative adversarial network-based model which can not only discriminate the realism of sDECT but also differentiate high-energy CT (HECT) and LECT from DECT. A cohort of 57 head-and-neck cancer patients with DECT and MRI pairs were used to validate the performance of the proposed framework. The results of sDECT and its derived SPR maps were compared with clinical DECT and the corresponding SPR, respectively. The mean absolute error for synthetic LECT and HECT were 79.98 ± 18.11 HU and 80.15 ± 16.27 HU, respectively. The corresponding SPR maps generated from sDECT showed a normalized mean absolute error as 5.22% ± 1.23%. By comparing with the traditional Cycle GANs, our proposed method significantly improves the accuracy of sDECT. The results indicate that on our dataset, the sDECT image form MRI is close to planning DECT, and thus shows promising potential for generating SPR maps for proton therapy.

Synthesizing CT images from MR images with deep learning: model generalization for different datasets through transfer learning

PURPOSE

Synthetic CT generation is the focus of many studies, however, only models on data with the same dataset were tested. Therefore, how well the trained model will work for data from different hospitals and MR protocols is still unknown. In this study, we addressed the model generalization problem for the MR-to-CT conversion task.

METHODS

Brain T2 MR and corresponding CT images were collected from one hospital and brain T1-FLAIR, T1-POST MR, and corresponding CT images were collected from another hospital. To investigate the model's generalizability ability, four potential solutions were proposed: source model, target model, combined model, and adapted model. All models were trained using the CycleGAN network. The source model was trained with a source domain dataset from scratch and tested with a target domain dataset. The target model was trained with a target domain dataset and tested with a target domain dataset. The combined model was trained with both source domain and target domain datasets, and tested with the target domain dataset. The adapted model used a transfer learning strategy to train a CycleGAN model with a source domain dataset and retrain the pre-trained model with a target domain dataset. MAE, RMSE, PSNR, and SSIM were used to quantitatively evaluate model performance on a target domain dataset.

RESULTS

The adapted model achieved best quantitative results of 74.56±8.61, 193.18±17.98, 28.30±0.83, and 0.84±0.01 for MAE, RMSE, PSNR, and SSIM using the T1-FLAIR dataset and 74.89±15.64, 195.73±31.29, 27.72±1.43, and 0.83±0.04 for MAE, RMSE, PSNR, and SSIM using the T1-POST dataset. The source model had the poorest performance.

CONCLUSIONS

This work indicates high generalization ability to generate synthetic CT images from small training datasets of MR images using pre-trained CycleGAN. The quantitative results of the test data, including different scanning protocols and different acquisition centers, indicated the proof of this concept.

An MR-only acquisition and artificial intelligence based image-processing protocol for photon and proton therapy using a low field MR

OBJECTIVE

Recent developments on synthetically generated CTs (sCT), hybrid MRI linacs and MR-only simulations underlined the clinical feasibility and acceptance of MR guided radiation therapy. However, considering clinical application of open and low field MR with a limited field of view can result in truncation of the patient's anatomy which further affects the MR to sCT conversion. In this study an acquisition protocol and subsequent MR image stitching is proposed to overcome the limited field of view restriction of open MR scanners, for MR-only photon and proton therapy.

MATERIAL AND METHODS

12 prostate cancer patients scanned with an open 0.35T scanner were included. To obtain the full body contour an enhanced imaging protocol including two repeated scans after bilateral table movement was introduced. All required structures (patient contour, target and organ at risk) were delineated on a post-processed combined transversal image set (stitched MRI). The postprocessed MR was converted into a sCT by a pretrained neural network generator. Inversely planned photon and proton plans (VMAT and SFUD) were designed using the sCT and recalculated for rigidly and deformably registered CT images and compared based on D2%, D50%, V70Gy for organs at risk and based on D2%, D50%, D98% for the CTV and PTV. The stitched MRI and the untruncated MRI were compared to the CT, and the maximum surface distance was calculated. The sCT was evaluated with respect to delineation accuracy by comparing on stitched MRI and sCT using the DICE coefficient for femoral bones and the whole body.

RESULTS

Maximum surface distance analysis revealed uncertainties in lateral direction of 1-3mm on average. DICE coefficient analysis confirms good performance of the sCT conversion, i.e. 92%, 93%, and 100% were obtained for femoral bone left and right and whole body. Dose comparison resulted in uncertainties below 1% between deformed CT and sCT and below 2% between rigidly registered CT and sCT in the CTV for photon and proton treatment plans.

DISCUSSION

A newly developed acquisition protocol for open MR scanners and subsequent Sct generation revealed good acceptance for photon and proton therapy. Moreover, this protocol tackles the restriction of the limited FOVs and expands the capacities towards MR guided proton therapy with horizontal beam lines.

Attenuation correction for human PET/MRI studies

Attenuation correction has been one of the main methodological challenges in the integrated positron emission tomography and magnetic resonance imaging (PET/MRI) field. As standard transmission or computed tomography approaches are not available in integrated PET/MRI scanners, MR-based attenuation correction approaches had to be developed. Aspects that have to be considered for implementing accurate methods include the need to account for attenuation in bone tissue, normal and pathological lung and the MR hardware present in the PET field-of-view, to reduce the impact of subject motion, to minimize truncation and susceptibility artifacts, and to address issues related to the data acquisition and processing both on the PET and MRI sides. The standard MR-based attenuation correction techniques implemented by the PET/MRI equipment manufacturers and their impact on clinical and research PET data interpretation and quantification are first discussed. Next, the more advanced methods, including the latest generation deep learning-based approaches that have been proposed for further minimizing the attenuation correction related bias are described. Finally, a future perspective focused on the needed developments in the field is given.

Dosimetric impact of geometric distortions in an MRI-only proton therapy workflow for lung, liver and pancreas

In a radiation therapy workflow based on Magnetic Resonance Imaging (MRI), dosimetric errors may arise due to geometric distortions introduced by MRI. The aim of this study was to quantify the dosimetric effect of system-dependent geometric distortions in an MRI-only workflow for proton therapy applied at extra-cranial sites. An approach was developed, in which computed tomography (CT) images were distorted using an MRI displacement map, which represented the MR distortions in a spoiled gradient-echo sequence due to gradient nonlinearities and static magnetic field inhomogeneities. A retrospective study was conducted on 4DCT/MRI digital phantoms and 18 4DCT clinical datasets of the thoraco-abdominal site. The treatment plans were designed and separately optimized for each beam in a beam specific Planning Target Volume on the distorted CT, and the final dose distribution was obtained as the average. The dose was then recalculated in undistorted CT using the same beam geometry and beam weights. The analysis was performed in terms of Dose Volume Histogram (DVH) parameters. No clinically relevant dosimetric impact was observed on organs at risk, whereas in the target structure, geometric distortions caused statistically significant variations in the planned dose DVH parameters and dose homogeneity index (DHI). The dosimetric variations in the target structure were smaller in abdominal cases (ΔD_2%, ΔD_98%, and ΔD_mean all below 0.1% and ΔDHI below 0.003) compared to the lung cases. Indeed, lung patients with tumors isolated inside lung parenchyma exhibited higher dosimetric variations (ΔD_2%≥0.3%, ΔD_98%≥15.9%, ΔD_mean≥3.3% and ΔDHI≥0.102) than lung patients with tumor close to soft tissue (ΔD_2%≤0.4%, ΔD_98%≤5.6%, ΔD_mean≤0.9% and ΔDHI≤0.027) potentially due to higher density variations along the beam path. Results suggest the potential applicability of MRI-only proton therapy, provided that specific analysis is applied for isolated lung tumors.

Self-derived organ attention for unpaired CT-MRI deep domain adaptation based MRI segmentation

To develop and evaluate a deep learning method to segment parotid glands from MRI using unannotated MRI and unpaired expert-segmented CT datasets. We introduced a new self-derived organ attention deep learning network for combined CT to MRI image-to-image translation (I2I) and MRI segmentation, all trained as an end-to-end network. The expert segmentations available on CT scans were combined with the I2I translated pseudo MR images to train the MRI segmentation network. Once trained, the MRI segmentation network alone is required. We introduced an organ attention discriminator that constrains the CT to MR generator to synthesize pseudo MR images that preserve organ geometry and appearance statistics as in real MRI. The I2I translation network training was regularized using the organ attention discriminator, global image-matching discriminator, and cycle consistency losses. MRI segmentation training was regularized by using cross-entropy loss. Segmentation performance was compared against multiple domain adaptation-based segmentation methods using the Dice similarity coefficient (DSC) and Hausdorff distance at the 95th percentile (HD95). All networks were trained using 85 unlabeled T2-weighted fat suppressed (T2wFS) MRIs and 96 expert-segmented CT scans. Performance upper-limit was based on fully supervised MRI training done using the 85 T2wFS MRI with expert segmentations. Independent evaluation was performed on 77 MRIs never used in training. The proposed approach achieved the highest accuracy (left parotid: DSC 0.82 ± 0.03, HD95 2.98 ± 1.01 mm; right parotid: 0.81 ± 0.05, HD95 3.14 ± 1.17 mm) compared to other methods. This accuracy was close to the reference fully supervised MRI segmentation (DSC of 0.84 ± 0.04, a HD95 of 2.24 ± 0.77 mm for the left parotid, and a DSC of 0.84 ± 0.06 and HD95 of 2.32 ± 1.37 mm for the right parotid glands).

Overview of artificial intelligence-based applications in radiotherapy: Recommendations for implementation and quality assurance

Artificial Intelligence (AI) is currently being introduced into different domains, including medicine. Specifically in radiation oncology, machine learning models allow automation and optimization of the workflow. A lack of knowledge and interpretation of these AI models can hold back wide-spread and full deployment into clinical practice. To facilitate the integration of AI models in the radiotherapy workflow, generally applicable recommendations on implementation and quality assurance (QA) of AI models are presented. For commonly used applications in radiotherapy such as auto-segmentation, automated treatment planning and synthetic computed tomography (sCT) the basic concepts are discussed in depth. Emphasis is put on the commissioning, implementation and case-specific and routine QA of AI models needed for a methodical introduction in clinical practice.