Technical Note: Dose prediction for radiation therapy using feature-based losses and One Cycle Learning

An Efficient 3D Convolutional Neural Network for Dose Prediction in Cancer Radiotherapy from CT Images

Introduction: Cancer is a highly lethal disease with a significantly high mortality rate. One of the most commonly used methods for treatment is radiation therapy. However, cancer treatment using radiotherapy is a time-consuming process that requires significant manual work from planners and doctors. In radiation therapy treatment planning, determining the dose distribution for each of the regions of the patient's body is one of the most difficult and important tasks. Nowadays, artificial intelligence has shown promising results in improving the quality of disease treatment, particularly in cancer radiation therapy. Objectives: The main objective of this study is to build a high-performance deep learning model for predicting radiation therapy doses for cancer and to develop software to easily manipulate and use this model. Materials and Methods: In this paper, we propose a custom 3D convolutional neural network model with a U-Net-based architecture to automatically predict radiation doses during cancer radiation therapy from CT images. To ensure that the predicted doses do not have negative values, which are not valid for radiation doses, a rectified linear unit (ReLU) function is applied to the output to convert negative values to zero. Additionally, a proposed loss function based on a dose-volume histogram is used to train the model, ensuring that the predicted dose concentrations are highly meaningful in terms of radiation therapy. The model is developed using the OpenKBP challenge dataset, which consists of 200, 100, and 40 head and neck cancer patients for training, testing, and validation, respectively. Before the training phase, preprocessing and augmentation techniques, such as standardization, translation, and flipping, are applied to the training set. During the training phase, a cosine annealing scheduler is applied to update the learning rate. Results and Conclusions: Our model achieved strong performance, with a good DVH score (1.444 Gy) on the test dataset, compared to previous studies and state-of-the-art models. In addition, we developed software to display the dose maps predicted by the proposed model for each 2D slice in order to facilitate usage and observation. These results may help doctors in treating cancer with radiation therapy in terms of both time and effectiveness.

Enhancing stereotactic ablative boost radiotherapy dose prediction for bulky lung cancer: A multi-scale dilated network approach with scale-balanced structure loss

PURPOSE

Partial stereotactic ablative boost radiotherapy (P-SABR) effectively treats bulky lung cancer; however, the planning process for P-SABR requires repeated dose calculations. To improve planning efficiency, we proposed a novel deep learning method that utilizes limited data to accurately predict the three-dimensional (3D) dose distribution of the P-SABR plan for bulky lung cancer.

METHODS

We utilized data on 74 patients diagnosed with bulky lung cancer who received P-SABR treatment. The patient dataset was randomly divided into a training set (51 plans) with augmentation, validation set (7 plans), and testing set (16 plans). We devised a 3D multi-scale dilated network (MD-Net) and integrated a scale-balanced structure loss into the loss function. A comparative analysis with a classical network and other advanced networks with multi-scale analysis capabilities and other loss functions was conducted based on the dose distributions in terms of the axial view, average dose scores (ADSs), and average absolute differences of dosimetric indices (AADDIs). Finally, we analyzed the predicted dosimetric indices against the ground-truth values and compared the predicted dose-volume histogram (DVH) with the ground-truth DVH.

RESULTS

Our proposed dose prediction method for P-SABR plans for bulky lung cancer demonstrated strong performance, exhibiting a significant improvement in predicting multiple indicators of regions of interest (ROIs), particularly the gross target volume (GTV). Our network demonstrated increased accuracy in most dosimetric indices and dose scores in different ROIs. The proposed loss function significantly enhanced the predictive performance of the dosimetric indices. The predicted dosimetric indices and DVHs were equivalent to the ground-truth values.

CONCLUSION

Our study presents an effective model based on limited datasets, and it exhibits high accuracy in the dose prediction of P-SABR plans for bulky lung cancer. This method has potential as an automated tool for P-SABR planning and can help optimize treatments and improve planning efficiency.

Ultra-fast, one-click radiotherapy treatment planning outside a treatment planning system

We present an automated radiation oncology treatment planning pipeline that operates between segmentation and plan review, minimizing manual interaction and reliance on traditional planning systems. Two AI models work in sequence: the first generates a dose distribution, and the second creates a deliverable DICOM-RT plan. Trained and validated on 276 plans, and tested on 151 datasets, the system produced clinically deliverable plans-complete with all VMAT parameters-in about 38 s. These plans met target coverage and most organ-at-risk constraints. This proof-of-concept demonstrates the feasibility of generating high-quality, deliverable DICOM plans within seconds.

Evaluation of deep learning based dose prediction in head and neck cancer patients using two different types of input contours

PURPOSE

This study evaluates deep learning (DL) based dose prediction methods in head and neck cancer (HNC) patients using two types of input contours.

MATERIALS AND METHODS

Seventy-five HNC patients undergoing two-step volumetric-modulated arc therapy were included. Dose prediction was performed using the AIVOT prototype (AiRato.Inc, Sendai, Japan), a commercial software with an HD U-net-based dose distribution prediction system. Models were developed for the initial plan (46 Gy/23Fr) and boost plan (24 Gy/12Fr), trained with 65 cases and tested with 10 cases. The 8-channel model used one target (PTV) and seven organs at risk (OARs), while the 10-channel model added two dummy contours (PTV ring and spinal cord PRV). Predicted and deliverable doses, obtained through dose mimicking on another radiation treatment planning system, were evaluated using dose-volume indices for PTV and OARs.

RESULTS

For the initial plan, both models achieved approximately 2% prediction accuracy for the target dose and maintained accuracy within 3.2 Gy for OARs. The 10-channel model outperformed the 8-channel model for certain dose indices. For the boost plan, both models exhibited prediction accuracies of approximately 2% for the target dose and 1 Gy for OARs. The 10-channel model showed significantly closer predictions to the ground truth for D50% and Dmean. Deliverable plans based on prediction doses showed little significant difference compared to the ground truth, especially for the boost plan.

CONCLUSION

DL-based dose prediction using the AIVOT prototype software in HNC patients yielded promising results. While additional contours may enhance prediction accuracy, their impact on dose mimicking is relatively small.

Dose prediction of CyberKnife Monte Carlo plan for lung cancer patients based on deep learning: robust learning of variable beam configurations

BACKGROUND

Accurate calculation of lung cancer dose using the Monte Carlo (MC) algorithm in CyberKnife (CK) is essential for precise planning. We aim to employ deep learning to directly predict the 3D dose distribution calculated by the MC algorithm, enabling rapid and accurate automatic planning. However, most current methods solely focus on conventional intensity-modulated radiation therapy and assume a consistent beam configuration across all patients. This study seeks to develop a more versatile model incorporating variable beam configurations of CK and considering the patient's anatomy.

METHODS

This study proposed that the AB (anatomy and beam) model be compared with the control Mask (only anatomy) model. These models are based on a 3D U-Net network to investigate the impact of CK beam encoding information on dose prediction. The study collected 86 lung cancer patients who received CK's built-in MC algorithm plans using different beam configurations for training/validation (66 cases) and testing (20 cases). We compared the gamma passing rate, dose difference maps, and relevant dose-volume metrics to evaluate the model's performance. In addition, the Dice similarity coefficients (DSCs) were calculated to assess the spatial correspondence of isodose volumes.

RESULTS

The AB model demonstrated superior performance compared to the Mask model, particularly in the trajectory dose of the beam. The DSCs of the AB model were 20-40% higher than that of the Mask model in some dose regions. We achieved approximately 99% for the PTV and generally more than 95% for the organs at risk (OARs) referred to the clinical planning dose in the gamma passing rates (3 mm/3%). Relative to the Mask model, the AB model exhibited more than 90% improvement in small voxels (p < 0.001). The AB model matched well with the clinical plan's dose-volume histograms, and the average dose error for all organs was 1.65 ± 0.69%.

CONCLUSIONS

Our proposed new model signifies a crucial advancement in predicting CK 3D dose distributions for clinical applications. It enables planners to rapidly and precisely predict MC doses for lung cancer based on patient-specific beam configurations and optimize the CK treatment process.

Automation of ePROMs in radiation oncology and its impact on patient response and bias

PURPOSE

This study evaluates the impact of integrating a novel, in-house developed electronic Patient-Reported Outcome Measures (ePROMs) tool with a commercial Oncology Information System (OIS) on patient response rates and potential biases in real-world data science applications.

MATERIALS AND METHODS

We designed an ePROMs tool using the NodeJS web application framework, automatically sending e-mail questionnaires to patients based on their treatment schedules in the OIS. The tool is used across various treatment sites to collect PROMs data in a real-world setting. This research examined the effects of increasing automation levels on both recruitment and response rates, as well as potential biases across different patient cohorts. Automation was implemented in three escalating levels, from telephone reminders for missing reports to minimal intervention from study nurses.

RESULTS

From August 2020 to December 2023, 1,944 patients participated in the PROMs study. Our findings indicate that automating the workflows substantially reduced the patient management workload. However, higher levels of automation led to lower response rates, particularly in collecting late-phase symptoms in breast and head-and-neck cancer cohorts. Additionally, email-based PROMs introduced an age bias when recruiting new patients for the ePROMs study. Nevertheless, age was not a significant predictor of early dropout or missing symptom reports among patients participating. Notably, increased automation was significantly correlated with lower response rates in breast (p = 0.026) and head-and-neck cancer patients (p < 0.001).

CONCLUSION

Integrating ePROMs within the OIS can significantly reduce workload and personnel resources. However, this efficiency may compromise patient responses in certain groups. A balance must be achieved between workload, resource allocation, and the sensitivity needed to detect clinically significant effects. This may necessitate customized automation levels tailored to specific cancer groups, highlighting a fundamental trade-off between operational efficiency and data quality.

Prediction of dose distributions for non-small cell lung cancer patients using MHA-ResUNet

BACKGROUND

The current level of automation in the production of radiotherapy plans for lung cancer patients is relatively low. With the development of artificial intelligence, it has become a reality to use neural networks to predict dose distributions and provide assistance for radiation therapy planning. However, due to the significant individual variability in the distribution of non-small cell lung cancer (NSCLC) planning target volume (PTV) and the complex spatial relationships between the PTV and organs at risk (OARs), there is still a lack of a high-precision dose prediction network tailored to the characteristics of NSCLC.

PURPOSE

To assist in the development of volumetric modulated arc therapy (VMAT) plans for non-small cell lung cancer patients, a deep neural network is proposed to predict high-precision dose distribution.

METHODS

This study has developed a network called MHA-ResUNet, which combines a large-kernel dilated convolution module and multi-head attention (MHA) modules. The network was trained based on 80 VMAT plans of NSCLC patients. CT images, PTV, and OARs were fed into the independent input channel. The dose distribution was taken as the output to train the model. The performance of this network was compared with that of several commonly used networks, and the networks' performance was evaluated based on the voxel-level mean absolute error (MAE) within the PTV and OARs, as well as the error in clinical dose-volume metrics.

RESULTS

The MAE between the predicted dose distribution and the manually planned dose distribution within the PTV is 1.43 Gy, and the D95 error is less than 1 Gy. Compared with the other three commonly used networks, the dose error of the MHA-ResUNet is the smallest in PTV and OARs.

CONCLUSIONS

The proposed MHA-ResUNet network improves the receptive field and filters the shallow features to learn the relative spatial relation between the PTV and the OARs, enabling accurate prediction of dose distributions in NSCLC patients undergoing VMAT radiotherapy.

Dosimetric Impact of Delineation and Motion Uncertainties on the Heart and Substructures in Lung Cancer Radiotherapy

AIMS

Delineation variations and organ motion produce difficult-to-quantify uncertainties in planned radiation doses to targets and organs at risk. Similar to manual contouring, most automatic segmentation tools generate single delineations per structure; however, this does not indicate the range of clinically acceptable delineations. This study develops a method to generate a range of automatic cardiac structure segmentations, incorporating motion and delineation uncertainty, and evaluates the dosimetric impact in lung cancer.

MATERIALS AND METHODS

Eighteen cardiac structures were delineated using a locally developed auto-segmentation tool. It was applied to lung cancer planning CTs for 27 curative (planned dose ≥50 Gy) cases, and delineation variations were estimated by using ten mapping-atlases to provide separate substructure segmentations. Motion-related cardiac segmentation variations were estimated by auto-contouring structures on ten respiratory phases for 9/27 cases that had 4D-planning CTs. Dose volume histograms (DVHs) incorporating these variations were generated for comparison.

RESULTS

Variations in mean doses (Dmean), defined as the range in values across ten feasible auto-segmentations, were calculated for each cardiac substructure. Over the study cohort the median variations for delineation uncertainty and motion were 2.20-11.09 Gy and 0.72-4.06 Gy, respectively. As relative values, variations in Dmean were between 18.7%-65.3% and 7.8%-32.5% for delineation uncertainty and motion, respectively. Doses vary depending on the individual planned dose distribution, not simply on segmentation differences, with larger dose variations to cardiac structures lying within areas of steep dose gradient.

CONCLUSION

Radiotherapy dose uncertainties from delineation variations and respiratory-related heart motion were quantified using a cardiac substructure automatic segmentation tool. This predicts the 'dose range' where doses to structures are most likely to fall, rather than single DVH curves. This enables consideration of these uncertainties in cardiotoxicity research and for future plan optimisation. The tool was designed for cardiac structures, but similar methods are potentially applicable to other OARs.

Multi-center Dose Prediction Using Attention-aware Deep learning Algorithm Based on Transformers for Cervical Cancer Radiotherapy

AIMS

Accurate dose delivery is crucial for cervical cancer volumetric modulated arc therapy (VMAT). We aimed to develop a robust deep-learning (DL) algorithm for fast and accurate dose prediction of cervical cancer VMAT in multicenter datasets and then explore the feasibility of the DL algorithm to endometrial cancer VMAT with different prescriptions.

MATERIALS AND METHODS

We proposed the AtTranNet algorithm for three-dimensional dose prediction. A total of 367 cervical patients were enrolled in this study. Three hundred twenty-two cervical patients from 3 centers were randomly divided into 70%, 10%, and 20% as training, validation, and testing sets, respectively. Forty-five cervical patients from another center were selected for external testing. Moreover, 70 patients of endometrial cancer with different prescriptions were further selected to test the model. Prediction precision was evaluated by dosimetric difference, dose map, and dose-volume histogram metrics.

RESULTS

The prediction results were all clinically acceptable. The mean absolute error within the body in internal testing was 0.66 ± 0.63%. The maximum |δD| for planning target volume was observed in D98, which is 1.24 ± 2.73 Gy. The maximum |δD| for organs at risk was observed in Dmean of bladder, which is 4.79 ± 3.14 Gy. The maximum |δV| were observed in V40 of pelvic bones, which is 4.77 ± 4.48%.

CONCLUSION

AtTranNet showed the feasibility and reasonable accuracy in the dose prediction for cervical cancer in multiple centers. The model can also be generalized for endometrial cancer with different prescriptions without any transfer learning.

Beam-wise dose composition learning for head and neck cancer dose prediction in radiotherapy

Automatic and accurate dose distribution prediction plays an important role in radiotherapy plan. Although previous methods can provide promising performance, most methods did not consider beam-shaped radiation of treatment delivery in clinical practice. This leads to inaccurate prediction, especially on beam paths. To solve this problem, we propose a beam-wise dose composition learning (BDCL) method for dose prediction in the context of head and neck (H&N) radiotherapy plan. Specifically, a global dose network is first utilized to predict coarse dose values in the whole-image space. Then, we propose to generate individual beam masks to decompose the coarse dose distribution into multiple field doses, called beam voters, which are further refined by a subsequent beam dose network and reassembled to form the final dose distribution. In particular, we design an overlap consistency module to keep the similarity of high-level features in overlapping regions between different beam voters. To make the predicted dose distribution more consistent with the real radiotherapy plan, we also propose a dose-volume histogram (DVH) calibration process to facilitate feature learning in some clinically concerned regions. We further apply an edge enhancement procedure to enhance the learning of the extracted feature from the dose falloff regions. Experimental results on a public H&N cancer dataset from the AAPM OpenKBP challenge show that our method achieves superior performance over other state-of-the-art approaches by significant margins. Source code is released at https://github.com/TL9792/BDCLDosePrediction.

Comparison of Ethos template-based planning and AI-based dose prediction: General performance, patient optimality, and limitations

PURPOSE

Ethos proposes a template-based automatic dose planning (Etb) for online adaptive radiotherapy. This study evaluates the general performance of Etb for prostate cancer, as well as the ability to generate patient-optimal plans, by comparing it with another state-of-the-art automatic planning method, i.e., deep learning dose prediction followed by dose mimicking (DP + DM).

MATERIALS

General performances and capability to produce patient-optimal plan were investigated through two studies: Study-S1 generated plans for 45 patients using our initial Ethos clinical goals template (EG_init), and compared them to manually generated plans (MG). For study-S2, 10 patients which showed poor performances at study-S1 were selected. S2 compared the quality of plans generated with four different methods: 1) Ethos initial template (EG_init_selected), 2) Ethos updated template-based on S1 results (EG_upd_selected), 3) DP + DM, and 4) MG plans.

RESULTS

EG_init plans showed satisfactory performance for dose level above 50 Gy: reported mean metrics differences (EG_init minus MG) never exceeded 0.6 %. However, lower dose levels showed loosely optimized metrics, mean differences for V30Gy to rectum and V20Gy to anal canal were of 6.6 % and 13.0 %. EG_init_selected showed amplified differences in V30Gy to rectum and V20Gy to anal canal: 8.5 % and 16.9 %, respectively. These dropped to 5.7 % and 11.5 % for EG_upd_selected plans but strongly increased V60Gy to rectum for 2 patients. DP + DM plans achieved differences of 3.4 % and 4.6 % without compromising any V60Gy.

CONCLUSION

General performances of Etb were satisfactory. However, optimizing with template of goals might be limiting for some complex cases. Over our test patients, DP + DM outperformed the Etb approach.

Deep learning-based dose prediction to improve the plan quality of volumetric modulated arc therapy for gynecologic cancers

BACKGROUND

In recent years, deep-learning models have been used to predict entire three-dimensional dose distributions. However, the usability of dose predictions to improve plan quality should be further investigated.

PURPOSE

To develop a deep-learning model to predict high-quality dose distributions for volumetric modulated arc therapy (VMAT) plans for patients with gynecologic cancer and to evaluate their usability in driving plan quality improvements.

METHODS

A total of 79 VMAT plans for the female pelvis were used to train (47 plans), validate (16 plans), and test (16 plans) 3D dense dilated U-Net models to predict 3D dose distributions. The models received the normalized CT scan, dose prescription, and target and normal tissue contours as inputs. Three models were used to predict the dose distributions for plans in the test set. A radiation oncologist specializing in the treatment of gynecologic cancers scored the test set predictions using a 5-point scale (5, acceptable as-is; 4, prefer minor edits; 3, minor edits needed; 2, major edits needed; and 1, unacceptable). The clinical plans for which the dose predictions indicated that improvements could be made were reoptimized with constraints extracted from the predictions.

RESULTS

The predicted dose distributions in the test set were of comparable quality to the clinical plans. The mean voxel-wise dose difference was -0.14 ± 0.46 Gy. The percentage dose differences in the predicted target metrics ofD 1 % ${D}_{1{\mathrm{\% }}}$ andD 98 % ${D}_{98{\mathrm{\% }}}$ were -1.05% ± 0.59% and 0.21% ± 0.28%, respectively. The dose differences in the predicted organ at risk mean and maximum doses were -0.30 ± 1.66 Gy and -0.42 ± 2.07 Gy, respectively. A radiation oncologist deemed all of the predicted dose distributions clinically acceptable; 12 received a score of 5, and four received a score of 4. Replanning of flagged plans (five plans) showed that the original plans could be further optimized to give dose distributions close to the predicted dose distributions.

CONCLUSIONS

Deep-learning dose prediction can be used to predict high-quality and clinically acceptable dose distributions for VMAT female pelvis plans, which can then be used to identify plans that can be improved with additional optimization.

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

Background and purpose

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

Materials and Methods

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

Results

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

Conclusion

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

A comparative study of deep learning-based knowledge-based planning methods for 3D dose distribution prediction of head and neck

PURPOSE

In this paper, we compare four novel knowledge-based planning (KBP) algorithms using deep learning to predict three-dimensional (3D) dose distributions of head and neck plans using the same patients' dataset and quantitative assessment metrics.

METHODS

A dataset of 340 oropharyngeal cancer patients treated with intensity-modulated radiation therapy was used in this study, which represents the AAPM OpenKBP - 2020 Grand Challenge dataset. Four 3D convolutional neural network architectures were built. The models were trained on 64% of the data set and validated on 16% for voxel-wise dose predictions: U-Net, attention U-Net, residual U-Net (Res U-Net), and attention Res U-Net. The trained models were then evaluated for their performance on a test data set (20% of the data) by comparing the predicted dose distributions against the ground-truth using dose statistics and dose-volume indices.

RESULTS

The four KBP dose prediction models exhibited promising performance with an averaged mean absolute dose error within the body contour <3 Gy on 68 plans in the test set. The average difference in predicting the D99 index for all targets was 0.92 Gy (p = 0.51) for attention Res U-Net, 0.94 Gy (p = 0.40) for Res U-Net, 2.94 Gy (p = 0.09) for attention U-Net, and 3.51 Gy (p = 0.08) for U-Net. For the OARs, the values for theD m a x ${D_{max}}$ andD m e a n ${D_{mean}}$ indices were 2.72 Gy (p < 0.01) for attention Res U-Net, 2.94 Gy (p < 0.01) for Res U-Net, 1.10 Gy (p < 0.01) for attention U-Net, 0.84 Gy (p < 0.29) for U-Net.

CONCLUSION

All models demonstrated almost comparable performance for voxel-wise dose prediction. KBP models that employ 3D U-Net architecture as a base could be deployed for clinical use to improve cancer patient treatment by creating plans with consistent quality and making the radiotherapy workflow more efficient.

Generating deliverable DICOM RT treatment plans for prostate VMAT by predicting MLC motion sequences with an encoder-decoder network

BACKGROUND

Deep learning-based auto-planning is an active research field; however, for some tasks a treatment planning system (TPS) is still required.

PURPOSE

To introduce a deep learning-based model generating deliverable DICOM RT treatment plans that can be directly irradiated by a linear accelerator (LINAC). The model was based on an encoder-decoder network and can predict multileaf collimator (MLC) motion sequences for prostate VMAT radiotherapy.

METHODS

A total of 619 treatment plans from 460 patients treated for prostate cancer with single-arc VMAT were included in this study. An encoder-decoder network was trained using 465 clinical treatment plans and validated on 77 plans. The performance was analyzed on a separate test set of 77 treatment plans. Separate L1 losses were computed for the leaf and jaw positions as well as the monitor units, with the leaf loss being weighted by a factor of 100 before being added to the other losses. The generated treatment plans were recalculated in a treatment planning system and the dose-volume metrics and gamma passing rates were compared to the original dose.

RESULTS

All generated treatment plans showed good agreement with the original data, with an average gamma passing rate (3%/3 mm) of 91.9 ± 7.1%. However, the coverage of the PTVs. was slightly lower for the generated plans (D98%  = 92.9 ± 2.6%) in comparison to the original plans (D98%  = 95.7 ± 2.2%). There was no significant difference in mean dose to the bladder between the predicted and original plan (Dmean of 28.0 ± 13.5 vs. 28.1 ± 13.3% of prescribed dose) or rectum (Dmean of 42.3 ± 7.4 vs. 42.6 ± 7.5%). The maximum dose to bladder was only slightly higher in the predicted plans (D2% of 100.7 ± 5.3 vs. 99.8 ± 4.0%) and for the rectum it was even lower (D2% of 100.5 ± 3.7 vs. 100.1 ± 4.3).

CONCLUSIONS

The deep learning-based model could predict MLC motion sequences in prostate VMAT plans, eliminating the need for sequencing inside a TPS, thus revolutionizing autonomous treatment planning workflows. This research completes the loop in deep learning-based treatment planning processes, enabling more efficient workflows for real-time or online adaptive radiotherapy.

Increasing Quality and Efficiency of the Radiotherapy Treatment Planning Process by Constructing and Implementing a Workflow-Monitoring Application

PURPOSE

To demonstrate how the efficiency of the treatment planning processes of a university radiation oncology department (2,500 new patients/year) could be improved by constructing and implementing a workflow-monitoring application.

METHODS

A web-based application was developed in house, which enhanced the process management tools of the clinic's oncology information system. The application calculates the days left for the next task in the treatment planning process and visualizes the information on a browser-based whiteboard. Workflow monitoring considers tumor types (breast, prostate, lung, etc) and treatment techniques and is backward planned from the planned start of treatment. The effect of introducing this application was analyzed over four phases: (1) baseline data without the workflow-monitoring application, (2) after introducing workflow visualization via a browser-based whiteboard, (3) after upgrading the whiteboard and introducing backend rules, and (4) after updating these rules on the basis of data from the previous phase.

RESULTS

Implementing the workflow-monitoring application and the introduced measures significantly reduced delays and, consequently, stress and a negative working atmosphere in the treatment planning process. Most notably, the amount of last-minute physics checks (on the day of the treatment start) could be reduced by 50%.

CONCLUSION

The study showed what measures can help organize and prioritize the treatment planning workflow. The increased efficiency is believed to improve the quality and reduce the risk of human error.

TrDosePred: A deep learning dose prediction algorithm based on transformers for head and neck cancer radiotherapy

BACKGROUND

Intensity-Modulated Radiation Therapy (IMRT) has been the standard of care for many types of tumors. However, treatment planning for IMRT is a time-consuming and labor-intensive process.

PURPOSE

To alleviate this tedious planning process, a novel deep learning based dose prediction algorithm (TrDosePred) was developed for head and neck cancers.

METHODS

The proposed TrDosePred, which generated the dose distribution from a contoured CT image, was a U-shape network constructed with a convolutional patch embedding and several local self-attention based transformers. Data augmentation and ensemble approach were used for further improvement. It was trained based on the dataset from Open Knowledge-Based Planning Challenge (OpenKBP). The performance of TrDosePred was evaluated with two mean absolute error (MAE) based scores utilized by OpenKBP challenge (i.e., Dose score and DVH score) and compared to the top three approaches of the challenge. In addition, several state-of-the-art methods were implemented and compared to TrDosePred.

RESULTS

The TrDosePred ensemble achieved the dose score of 2.426 Gy and the DVH score of 1.592 Gy on the test dataset, ranking at 3rd and 9th respectively in the leaderboard on CodaLab as of writing. In terms of DVH metrics, on average, the relative MAE against the clinical plans was 2.25% for targets and 2.17% for organs at risk.

CONCLUSIONS

A transformer-based framework TrDosePred was developed for dose prediction. The results showed a comparable or superior performance as compared to the previous state-of-the-art approaches, demonstrating the potential of transformer to boost the treatment planning procedures.

Deep Learning-Based Dose Prediction for Automated, Individualized Quality Assurance of Head and Neck Radiation Therapy Plans

PURPOSE

This study aimed to use deep learning-based dose prediction to assess head and neck (HN) plan quality and identify suboptimal plans.

METHODS AND MATERIALS

A total of 245 volumetric modulated arc therapy HN plans were created using RapidPlan knowledge-based planning (KBP). A subset of 112 high-quality plans was selected under the supervision of an HN radiation oncologist. We trained a 3D Dense Dilated U-Net architecture to predict 3-dimensional dose distributions using 3-fold cross-validation on 90 plans. Model inputs included computed tomography images, target prescriptions, and contours for targets and organs at risk (OARs). The model's performance was assessed on the remaining 22 test plans. We then tested the application of the dose prediction model for automated review of plan quality. Dose distributions were predicted on 14 clinical plans. The predicted versus clinical OAR dose metrics were compared to flag OARs with suboptimal normal tissue sparing using a 2 Gy dose difference or 3% dose-volume threshold. OAR flags were compared with manual flags by 3 HN radiation oncologists.

RESULTS

The predicted dose distributions were of comparable quality to the KBP plans. The differences between the predicted and KBP-planned D1%,D95%, and D99% across the targets were within -2.53% ± 1.34%, -0.42% ± 1.27%, and -0.12% ± 1.97%, respectively, and the OAR mean and maximum doses were within -0.33 ± 1.40 Gy and -0.96 ± 2.08 Gy, respectively. For the plan quality assessment study, radiation oncologists flagged 47 OARs for possible plan improvement. There was high interphysician variability; 83% of physician-flagged OARs were flagged by only one of 3 physicians. The comparative dose prediction model flagged 63 OARs, including 30 of 47 physician-flagged OARs.

CONCLUSIONS

Deep learning can predict high-quality dose distributions, which can be used as comparative dose distributions for automated, individualized assessment of HN plan quality.

OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines

Objective.To establish an open framework for developing plan optimization models for knowledge-based planning (KBP).Approach.Our framework includes radiotherapy treatment data (i.e. reference plans) for 100 patients with head-and-neck cancer who were treated with intensity-modulated radiotherapy. That data also includes high-quality dose predictions from 19 KBP models that were developed by different research groups using out-of-sample data during the OpenKBP Grand Challenge. The dose predictions were input to four fluence-based dose mimicking models to form 76 unique KBP pipelines that generated 7600 plans (76 pipelines × 100 patients). The predictions and KBP-generated plans were compared to the reference plans via: the dose score, which is the average mean absolute voxel-by-voxel difference in dose; the deviation in dose-volume histogram (DVH) points; and the frequency of clinical planning criteria satisfaction. We also performed a theoretical investigation to justify our dose mimicking models.Main results.The range in rank order correlation of the dose score between predictions and their KBP pipelines was 0.50-0.62, which indicates that the quality of the predictions was generally positively correlated with the quality of the plans. Additionally, compared to the input predictions, the KBP-generated plans performed significantly better (P< 0.05; one-sided Wilcoxon test) on 18 of 23 DVH points. Similarly, each optimization model generated plans that satisfied a higher percentage of criteria than the reference plans, which satisfied 3.5% more criteria than the set of all dose predictions. Lastly, our theoretical investigation demonstrated that the dose mimicking models generated plans that are also optimal for an inverse planning model.Significance.This was the largest international effort to date for evaluating the combination of KBP prediction and optimization models. We found that the best performing models significantly outperformed the reference dose and dose predictions. In the interest of reproducibility, our data and code is freely available.

Attention-aware 3D U-Net convolutional neural network for knowledge-based planning 3D dose distribution prediction of head-and-neck cancer

PURPOSE

Deep learning-based knowledge-based planning (KBP) methods have been introduced for radiotherapy dose distribution prediction to reduce the planning time and maintain consistent high-quality plans. This paper presents a novel KBP model using an attention-gating mechanism and a three-dimensional (3D) U-Net for intensity-modulated radiation therapy (IMRT) 3D dose distribution prediction in head-and-neck cancer.

METHODS

A total of 340 head-and-neck cancer plans, representing the OpenKBP-2020 AAPM Grand Challenge data set, were used in this study. All patients were treated with the IMRT technique and a dose prescription of 70 Gy. The data set was randomly divided into 64%/16%/20% as training/validation/testing cohorts. An attention-gated 3D U-Net architecture model was developed to predict full 3D dose distribution. The developed model was trained using the mean-squared error loss function, Adam optimization algorithm, a learning rate of 0.001, 120 epochs, and batch size of 4. In addition, a baseline U-Net model was also similarly trained for comparison. The model performance was evaluated on the testing data set by comparing the generated dose distributions against the ground-truth dose distributions using dose statistics and clinical dosimetric indices. Its performance was also compared to the baseline model and the reported results of other deep learning-based dose prediction models.

RESULTS

The proposed attention-gated 3D U-Net model showed high capability in accurately predicting 3D dose distributions that closely replicated the ground-truth dose distributions of 68 plans in the test set. The average value of the mean absolute dose error was 2.972 ± 1.220 Gy (vs. 2.920 ± 1.476 Gy for a baseline U-Net) in the brainstem, 4.243 ± 1.791 Gy (vs. 4.530 ± 2.295 Gy for a baseline U-Net) in the left parotid, 4.622 ± 1.975 Gy (vs. 4.223 ± 1.816 Gy for a baseline U-Net) in the right parotid, 3.346 ± 1.198 Gy (vs. 2.958 ± 0.888 Gy for a baseline U-Net) in the spinal cord, 6.582 ± 3.748 Gy (vs. 5.114 ± 2.098 Gy for a baseline U-Net) in the esophagus, 4.756 ± 1.560 Gy (vs. 4.992 ± 2.030 Gy for a baseline U-Net) in the mandible, 4.501 ± 1.784 Gy (vs. 4.925 ± 2.347 Gy for a baseline U-Net) in the larynx, 2.494 ± 0.953 Gy (vs. 2.648 ± 1.247 Gy for a baseline U-Net) in the PTV_70, and 2.432 ± 2.272 Gy (vs. 2.811 ± 2.896 Gy for a baseline U-Net) in the body contour. The average difference in predicting the D₉₉ value for the targets (PTV_70, PTV_63, and PTV_56) was 2.50 ± 1.77 Gy. For the organs at risk, the average difference in predicting the D m a x ${D_{max}}$ (brainstem, spinal cord, and mandible) and D m e a n ${D_{mean}}$ (left parotid, right parotid, esophagus, and larynx) values was 1.43 ± 1.01 and 2.44 ± 1.73 Gy, respectively. The average value of the homogeneity index was 7.99 ± 1.45 for the predicted plans versus 5.74 ± 2.95 for the ground-truth plans, whereas the average value of the conformity index was 0.63 ± 0.17 for the predicted plans versus 0.89 ± 0.19 for the ground-truth plans. The proposed model needs less than 5 s to predict a full 3D dose distribution of 64 × 64 × 64 voxels for a new patient that is sufficient for real-time applications.

CONCLUSIONS

The attention-gated 3D U-Net model demonstrated a capability in predicting accurate 3D dose distributions for head-and-neck IMRT plans with consistent quality. The prediction performance of the proposed model was overall superior to a baseline standard U-Net model, and it was also competitive to the performance of the best state-of-the-art dose prediction method reported in the literature. The proposed model could be used to obtain dose distributions for decision-making before planning, quality assurance of planning, and guiding-automated planning for improved plan consistency, quality, and planning efficiency.

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").

Dose Prediction Using a Three-Dimensional Convolutional Neural Network for Nasopharyngeal Carcinoma With Tomotherapy

Purpose

This study focused on predicting 3D dose distribution at high precision and generated the prediction methods for nasopharyngeal carcinoma patients (NPC) treated with Tomotherapy based on the patient-specific gap between organs at risk (OARs) and planning target volumes (PTVs).

Methods

A convolutional neural network (CNN) is trained using the CT and contour masks as the input and dose distributions as output. The CNN is based on the "3D Dense-U-Net", which combines the U-Net and the Dense-Net. To evaluate the model, we retrospectively used 124 NPC patients treated with Tomotherapy, in which 96 and 28 patients were randomly split and used for model training and test, respectively. We performed comparison studies using different training matrix shapes and dimensions for the CNN models, i.e., 128 ×128 ×48 (for Model I), 128 ×128 ×16 (for Model II), and 2D Dense U-Net (for Model III). The performance of these models was quantitatively evaluated using clinically relevant metrics and statistical analysis.

Results

We found a more considerable height of the training patch size yields a better model outcome. The study calculated the corresponding errors by comparing the predicted dose with the ground truth. The mean deviations from the mean and maximum doses of PTVs and OARs were 2.42 and 2.93%. Error for the maximum dose of right optic nerves in Model I was 4.87 ± 6.88%, compared with 7.9 ± 6.8% in Model II (p=0.08) and 13.85 ± 10.97% in Model III (p<0.01); the Model I performed the best. The gamma passing rates of PTV₆₀ for 3%/3 mm criteria was 83.6 ± 5.2% in Model I, compared with 75.9 ± 5.5% in Model II (p<0.001) and 77.2 ± 7.3% in Model III (p<0.01); the Model I also gave the best outcome. The prediction error of D₉₅ for PTV₆₀ was 0.64 ± 0.68% in Model I, compared with 2.04 ± 1.38% in Model II (p<0.01) and 1.05 ± 0.96% in Model III (p=0.01); the Model I was also the best one.

Conclusions

It is significant to train the dose prediction model by exploiting deep-learning techniques with various clinical logic concepts. Increasing the height (Y direction) of training patch size can improve the dose prediction accuracy of tiny OARs and the whole body. Our dose prediction network model provides a clinically acceptable result and a training strategy for a dose prediction model. It should be helpful to build automatic Tomotherapy planning.

OpenKBP: The open-access knowledge-based planning grand challenge and dataset

PURPOSE

To advance fair and consistent comparisons of dose prediction methods for knowledge-based planning (KBP) in radiation therapy research.

METHODS

We hosted OpenKBP, a 2020 AAPM Grand Challenge, and challenged participants to develop the best method for predicting the dose of contoured computed tomography (CT) images. The models were evaluated according to two separate scores: (a) dose score, which evaluates the full three-dimensional (3D) dose distributions, and (b) dose-volume histogram (DVH) score, which evaluates a set DVH metrics. We used these scores to quantify the quality of the models based on their out-of-sample predictions. To develop and test their models, participants were given the data of 340 patients who were treated for head-and-neck cancer with radiation therapy. The data were partitioned into training ( n = 200 ), validation ( n = 40 ), and testing ( n = 100 ) datasets. All participants performed training and validation with the corresponding datasets during the first (validation) phase of the Challenge. In the second (testing) phase, the participants used their model on the testing data to quantify the out-of-sample performance, which was hidden from participants and used to determine the final competition ranking. Participants also responded to a survey to summarize their models.

RESULTS

The Challenge attracted 195 participants from 28 countries, and 73 of those participants formed 44 teams in the validation phase, which received a total of 1750 submissions. The testing phase garnered submissions from 28 of those teams, which represents 28 unique prediction methods. On average, over the course of the validation phase, participants improved the dose and DVH scores of their models by a factor of 2.7 and 5.7, respectively. In the testing phase one model achieved the best dose score (2.429) and DVH score (1.478), which were both significantly better than the dose score (2.564) and the DVH score (1.529) that was achieved by the runner-up models. Lastly, many of the top performing teams reported that they used generalizable techniques (e.g., ensembles) to achieve higher performance than their competition.

CONCLUSION

OpenKBP is the first competition for knowledge-based planning research. The Challenge helped launch the first platform that enables researchers to compare KBP prediction methods fairly and consistently using a large open-source dataset and standardized metrics. OpenKBP has also democratized KBP research by making it accessible to everyone, which should help accelerate the progress of KBP research. The OpenKBP datasets are available publicly to help benchmark future KBP research.