Incorporating human and learned domain knowledge into training deep neural networks: A differentiable dose-volume histogram and adversarial inspired framework for generating Pareto optimal dose distributions in radiation therapy

LENAS: Learning-Based Neural Architecture Search and Ensemble for 3-D Radiotherapy Dose Prediction

Radiation therapy treatment planning requires balancing the delivery of the target dose while sparing normal tissues, making it a complex process. To streamline the planning process and enhance its quality, there is a growing demand for knowledge-based planning (KBP). Ensemble learning has shown impressive power in various deep learning tasks, and it has great potential to improve the performance of KBP. However, the effectiveness of ensemble learning heavily depends on the diversity and individual accuracy of the base learners. Moreover, the complexity of model ensembles is a major concern, as it requires maintaining multiple models during inference, leading to increased computational cost and storage overhead. In this study, we propose a novel learning-based ensemble approach named LENAS, which integrates neural architecture search with knowledge distillation for 3-D radiotherapy dose prediction. Our approach starts by exhaustively searching each block from an enormous architecture space to identify multiple architectures that exhibit promising performance and significant diversity. To mitigate the complexity introduced by the model ensemble, we adopt the teacher-student paradigm, leveraging the diverse outputs from multiple learned networks as supervisory signals to guide the training of the student network. Furthermore, to preserve high-level semantic information, we design a hybrid loss to optimize the student network, enabling it to recover the knowledge embedded within the teacher networks. The proposed method has been evaluated on two public datasets: 1) OpenKBP and 2) AIMIS. Extensive experimental results demonstrate the effectiveness of our method and its superior performance to the state-of-the-art methods. Code: github.com/hust-linyi/LENAS.

Deep reinforcement learning in radiation therapy planning optimization: A comprehensive review

PURPOSE

The formulation and optimization of radiation therapy plans are complex and time-consuming processes that heavily rely on the expertise of medical physicists. Consequently, there is an urgent need for automated optimization methods. Recent advancements in reinforcement learning, particularly deep reinforcement learning (DRL), show great promise for automating radiotherapy planning. This review summarizes the current state of DRL applications in this field, evaluates their effectiveness, and identifies challenges and future directions.

METHODS

A systematic search was conducted in Google Scholar, PubMed, IEEE Xplore, and Scopus using keywords such as "deep reinforcement learning", "radiation therapy", and "treatment planning". The extracted data were synthesized for an overview and critical analysis.

RESULTS

The application of deep reinforcement learning in radiation therapy plan optimization can generally be divided into three categories: optimizing treatment planning parameters, directly optimizing machine parameters, and adaptive radiotherapy. From the perspective of disease sites, DRL has been applied to cervical cancer, prostate cancer, vestibular schwannoma, and lung cancer. Regarding types of radiation therapy, it has been used in HDRBT, IMRT, SBRT, VMAT, GK, and Cyberknife.

CONCLUSIONS

Deep reinforcement learning technology has played a significant role in advancing the automated optimization of radiation therapy plans. However, there is still a considerable gap before it can be widely applied in clinical settings due to three main reasons: inefficiency, limited methods for quality assessment, and poor interpretability. To address these challenges, significant research opportunities exist in the future, such as constructing evaluators, parallelized training, and exploring continuous action spaces.

Enhancing neuro-oncology care through equity-driven applications of artificial intelligence

The disease course and clinical outcome for brain tumor patients depend not only on the molecular and histological features of the tumor but also on the patient's demographics and social determinants of health. While current investigations in neuro-oncology have broadly utilized artificial intelligence (AI) to enrich tumor diagnosis and more accurately predict treatment response, postoperative complications, and survival, equity-driven applications of AI have been limited. However, AI applications to advance health equity in the broader medical field have the potential to serve as practical blueprints to address known disparities in neuro-oncologic care. In this consensus review, we will describe current applications of AI in neuro-oncology, postulate viable AI solutions for the most pressing inequities in neuro-oncology based on broader literature, propose a framework for the effective integration of equity into AI-based neuro-oncology research, and close with the limitations of AI.

Neural network dose prediction for cervical brachytherapy: Overcoming data scarcity for applicator-specific models

BACKGROUND

3D neural network dose predictions are useful for automating brachytherapy (BT) treatment planning for cervical cancer. Cervical BT can be delivered with numerous applicators, which necessitates developing models that generalize to multiple applicator types. The variability and scarcity of data for any given applicator type poses challenges for deep learning.

PURPOSE

The goal of this work was to compare three methods of neural network training-a single model trained on all applicator data, fine-tuning the combined model to each applicator, and individual (IDV) applicator models-to determine the optimal method for dose prediction.

METHODS

Models were produced for four applicator types-tandem-and-ovoid (T&O), T&O with 1-7 needles (T&ON), tandem-and-ring (T&R) and T&R with 1-4 needles (T&RN). First, the combined model was trained on 859 treatment plans from 266 cervical cancer patients treated from 2010 onwards. The train/validation/test split was 70%/16%/14%, with approximately 49%/10%/19%/22% T&O/T&ON/T&R/T&RN in each dataset. Inputs included four channels for anatomical masks (high-risk clinical target volume [HRCTV], bladder, rectum, and sigmoid), a mask indicating dwell position locations, and applicator channels for each applicator component. Applicator channels were created by mapping the 3D dose for a single dwell position to each dwell position and summing over each applicator component with uniform dwell time weighting. A 3D Cascade U-Net, which consists of two U-Nets in sequence, and mean squared error loss function were used. The combined model was then fine-tuned to produce four applicator-specific models by freezing the first U-Net and encoding layers of the second and resuming training on applicator-specific data. Finally, four IDV models were trained using only data from each applicator type. Performance of these three model types was compared using the following metrics for the test set: mean error (ME, representing model bias) and mean absolute error (MAE) over all dose voxels and ME of clinical metrics (HRCTV D90% and D2cc of bladder, rectum, and sigmoid), averaged over all patients. A positive ME indicates the clinical dose was higher than predicted. 3D global gamma analysis with the prescription dose as reference value was performed. Dice similarity coefficients (DSC) were computed for each isodose volume.

RESULTS

Fine-tuned and combined models showed better performance than IDV applicator training. Fine-tuning resulted in modest improvements in about half the metrics, compared to the combined model, while the remainder were mostly unchanged. Fine-tuned MAE = 3.98%/2.69%/5.36%/3.80% for T&O/T&R/T&ON/T&RN, and ME over all voxels = -0.08%/-0.89%/-0.59%/1.42%. ME D2cc were bladder = -0.77%/1.00%/-0.66%/-1.53%, rectum = 1.11%/-0.22%/-0.29%/-3.37%, sigmoid = -0.47%/-0.06%/-2.37%/-1.40%, and ME D90 = 2.6%/-4.4%/4.8%/0.0%. Gamma pass rates (3%/3 mm) were 86%/91%/83%/89%. Mean DSCs were 0.92%/0.92%/0.88%/0.91% for isodoses ≤ 150% of prescription.

CONCLUSIONS

3D BT dose was accurately predicted for all applicator types, as indicated by the low MAE and MEs, high gamma scores and high DSCs. Training on all treatment data overcomes challenges with data scarcity in each applicator type, resulting in superior performance than can be achieved by training on IDV applicators alone. This could presumably be explained by the fact that the larger, more diverse dataset allows the neural network to learn underlying trends and characteristics in dose that are common to all treatment applicators. Accurate, applicator-specific dose predictions could enable automated, knowledge-based planning for any cervical brachytherapy treatment.

Prediction and visualization of moisture content in Tencha drying processes by computer vision and deep learning

BACKGROUND

It is important to monitor and control the moisture content throughout the Tencha drying processing procedure so that its quality is ensured. Workers often rely on their senses to perceive the moisture content, leading to relative subjectivity and low reproducibility. Traditional drying methods, which are used for measuring moisture content, are destructive to samples. This research was conducted using computer vision combined with deep learning to detect moisture content during the Tencha drying process. Different color space components of Tencha drying sample images were first extracted by computer vision. The color components were preprocessed using MinMax and Z score. Subsequently, one-dimensional convolutional neural networks (1D-CNN), partial least squares, and backpropagation artificial neural networks models were built and compared.

RESULTS

The 1D-CNN model and Z score preprocessing achieved superior predictive accuracy, with correlation coefficient of prediction (Rp) = 0.9548 for moisture content. The migration of moisture content during the Tencha drying process was eventually visualized by mapping its spatial and temporal distributions.

CONCLUSION

The results indicated that computer vision combined with 1D-CNN was feasible for moisture prediction during the Tencha drying process. This study provides technical support for the industrial and intelligent production of Tencha. © 2024 Society of Chemical Industry.

Clinical VMAT machine parameter optimization for localized prostate cancer using deep reinforcement learning

BACKGROUND

Volumetric modulated arc therapy (VMAT) machine parameter optimization (MPO) remains computationally expensive and sensitive to input dose objectives creating challenges for manual and automatic planning. Reinforcement learning (RL) involves machine learning through extensive trial-and-error, demonstrating performance exceeding humans, and existing algorithms in several domains.

PURPOSE

To develop and evaluate an RL approach for VMAT MPO for localized prostate cancer to rapidly and automatically generate deliverable VMAT plans for a clinical linear accelerator (linac) and compare resultant dosimetry to clinical plans.

METHODS

We extended our previous RL approach to enable VMAT MPO of a 3D beam model for a clinical linac through a policy network. It accepts an input state describing the current control point and predicts continuous machine parameters for the next control point, which are used to update the input state, repeating until plan termination. RL training was conducted to minimize a dose-based cost function for prescription of 60 Gy in 20 fractions using CT scans and contours from 136 retrospective localized prostate cancer patients, 20 of which had existing plans used to initialize training. Data augmentation was employed to mitigate over-fitting, and parameter exploration was achieved using Gaussian perturbations. Following training, RL VMAT was applied to an independent cohort of 15 patients, and the resultant dosimetry was compared to clinical plans. We also combined the RL approach with our clinical treatment planning system (TPS) to automate final plan refinement, and creating the potential for manual review and edits as required for clinical use.

RESULTS

RL training was conducted for 5000 iterations, producing 40 000 plans during exploration. Mean ± SD execution time to produce deliverable VMAT plans in the test cohort was 3.3 ± 0.5 s which were automatically refined in the TPS taking an additional 77.4 ± 5.8 s. When normalized to provide equivalent target coverage, the RL+TPS plans provided a similar mean ± SD overall maximum dose of 63.2 ± 0.6 Gy and a lower mean rectum dose of 17.4 ± 7.4 compared to 63.9 ± 1.5 Gy (p = 0.061) and 21.0 ± 6.0 (p = 0.024) for the clinical plans.

CONCLUSIONS

An approach for VMAT MPO using RL for a clinical linac model was developed and applied to automatically generate deliverable plans for localized prostate cancer patients, and when combined with the clinical TPS shows potential to rapidly generate high-quality plans. The RL VMAT approach shows promise to discover advanced linac control policies through trial-and-error, and algorithm limitations and future directions are identified and discussed.

Human-like intelligent automatic treatment planning of head and neck cancer radiation therapy

Objective.Automatic treatment planning of radiation therapy (RT) is desired to ensure plan quality, improve planning efficiency, and reduce human errors. We have proposed an Intelligent Automatic Treatment Planning framework with a virtual treatment planner (VTP), an artificial intelligence robot built using deep reinforcement learning, autonomously operating a treatment planning system (TPS). This study extends our previous successes in relatively simple prostate cancer RT planning to head-and-neck (H&N) cancer, a more challenging context even for human planners due to multiple prescription levels, proximity of targets to critical organs, and tight dosimetric constraints.Approach.We integrated VTP with a real clinical TPS to establish a fully automated planning workflow guided by VTP. This integration allowed direct model training and evaluation using the clinical TPS. We designed the VTP network structure to approach the decision-making process in RT planning in a hierarchical manner that mirrors human planners. The VTP network was trained via theQ-learning framework. To assess the effectiveness of VTP, we conducted a prospective evaluation in the 2023 Planning Challenge organized by the American Association of Medical Dosimetrists (AAMD). We extended our evaluation to include 20 clinical H&N cancer patients, comparing the plans generated by VTP against their clinical plans.Main results.In the prospective evaluation for the AAMD Planning Challenge, VTP achieved a plan score of 139.08 in the initial phase evaluating plan quality, and 15 min of planning time with the first place ranking in the adaptive phase competing for planning efficiency while meeting all plan quality requirements. For clinical cases, VTP-generated plans achieved an average VTP score of125.33±11.12, which outperformed the corresponding clinical plans with an average score of117.76±13.56.Significance.We successfully integrated VTP with the clinical TPS to achieve a fully automated treatment planning workflow. The compelling performance of VTP demonstrated its potential in automating H&N cancer RT planning.

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.

Advantages of IMRT optimization with MCO compared to IMRT optimization without MCO in reducing small bowel high dose index for cervical cancer patients-individualized treatment options

Background

Traditional intensity-modulated radiation therapy (IMRT) planning for cervical cancer is time-consuming and require iterative repeated optimization. In this study, we focused on leveraging multi-criteria optimization (MCO) to reduce the impact of small bowel high-dose indices on other optimization targets, thereby providing a rapid approach to individualized IMRT for cervical cancer patients.

Methods

Our research involved a cohort of 25 cervical cancer patients who underwent IMRT radiotherapy. The patient inclusion criteria were as follows: (I) histopathological confirmation of cervical cancer, (II) underwent IMRT radiation therapy, and (III) a prescribed dose of 180 cGy/28 fractions for the patient. All plans were replanned by an experienced dosimetrist without the MCO (W-IMRT). On the basis of the W-IMRT plan, the individualized IMRT (I-IMRT) plan was generated under the priority trade-off of reducing the D2cc (D2cc is the minimal dose to the 2 cm3 of the small bowel receiving the maximal dose) index of the small bowel using the MCO method, maintaining target coverage and protecting other organs at risk (OARs) as much as possible. Statistical analysis was performed using the Wilcoxon signature rank test.

Results

When the MCO method was applied to the IMRT plan, the high dose index decreased in the overlapping area between the small bowel and the planning treatment volume (PTV) (P<0.001, respectively). The D2cc index of the small bowel decreased to below 5,200 cGy in all I-IMRT plans. On the other hand, in PTV, the I-IMRT plan achieved a better homogeneity index (HI) compared to the W-IMRT plan. Significant dose reductions were also observed in the bladder (Dmean 144.8 cGy and V40 1.45%) (P<0.001, respectively), rectum (Dmean 43.9 cGy and V40 2.7%) (P<0.001, respectively) and bilateral femur heads (Dmean 150 cGy) (P<0.001, respectively).

Conclusions

Dosimetric differences suggest that the I-IMRT plan using the MCO method provides better protection of other OARs and equivalently in PTV coverage, while lowering the high-dose index in the small bowel as much as possible for patients with cervical cancer, thus providing a rapid approach to achieving individualized IMRT for cervical cancer patients.

Medical informed machine learning: A scoping review and future research directions

Combining domain knowledge (DK) and machine learning is a recent research stream to overcome multiple issues like limited explainability, lack of data, and insufficient robustness. Most approaches applying informed machine learning (IML), however, are customized to solve one specific problem. This study analyzes the status of IML in medicine by conducting a scoping literature review based on an existing taxonomy. We identified 177 papers and analyzed them regarding the used DK, the implemented machine learning model, and the motives for performing IML. We find an immense role of expert knowledge and image data in medical IML. We then provide an overview and analysis of recent approaches and supply five directions for future research. This review can help develop future medical IML approaches by easily referencing existing solutions and shaping future research directions.

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.

Implementation and evaluation of an intelligent automatic treatment planning robot for prostate cancer stereotactic body radiation therapy

PURPOSE

We previously developed a virtual treatment planner (VTP), an artificial intelligence robot, operating a treatment planning system (TPS). Using deep reinforcement learning guided by human knowledge, we trained the VTP to autonomously adjust relevant parameters in treatment plan optimization, similar to a human planner, to generate high-quality plans for prostate cancer stereotactic body radiation therapy (SBRT). This study describes the clinical implementation and evaluation of VTP.

MATERIALS AND METHODS

We integrate VTP with Eclipse TPS using scripting Application Programming Interface. VTP observes dose-volume histograms of relevant structures, decides how to adjust dosimetric constraints, including doses, volumes, and weighting factors, and applies the adjustments to the TPS interface to launch the optimization engine. This process continues until a high-quality plan is achieved. We evaluated VTP's performance using the prostate SBRT case from the 2016 American Association of Medical Dosimetrist/Radiosurgery Society plan study with its plan scoring system, and compared to human-generated plans submitted to the challenge. Using the same scoring system, we also compared the plan quality of 36 prostate SBRT cases (20 planned with IMRT and 16 planned with VMAT) treated at our institution for both VTP and human-generated plans.

RESULTS

In the plan study case, VTP achieved a score of 142.1/150.0, ranking the third in the competition (median 134.6). For the clinical cases, VTP achieved 110.6 ± 6.5 for 20 IMRT plans and 126.2 ± 4.7 for 16 VMAT plans, similar to scores of human-generated plans with 110.4 ± 7.0 for IMRT plans and 125.4 ± 4.4 for VMAT plans. The workflow, plan quality and planning time of VTP were reviewed to be satisfactory by experienced physicists.

CONCLUSION

We successfully implemented VTP to operate a TPS for autonomous human-like treatment planning for prostate SBRT.

Automated treatment planning framework for brachytherapy of cervical cancer using 3D dose predictions

Objective. To lay the foundation for automated knowledge-based brachytherapy treatment planning using 3D dose estimations, we describe an optimization framework to convert brachytherapy dose distributions directly into dwell times (DTs).Approach. A dose rate kernelḋ(r,θ,φ)was produced by exporting 3D dose for one dwell position from the treatment planning system and normalizing by DT. By translating and rotating this kernel to each dwell position, scaling by DT and summing over all dwell positions, dose was computed (Dcalc). We used a Python-coded COBYLA optimizer to iteratively determine the DTs that minimize the mean squared error betweenDcalcand reference doseDref, computed using voxels withDref80%-120% of prescription. As validation of the optimization, we showed that the optimizer replicates clinical plans whenDref= clinical dose in 40 patients treated with tandem-and-ovoid (T&O) or tandem-and-ring (T&R) and 0-3 needles. Then we demonstrated automated planning in 10 T&O usingDref= dose predicted from a convolutional neural network developed in past work. Validation and automated plans were compared to clinical plans using mean absolute differences (MAD=1N∑n=1Nabsxn-xn') over all voxels (xn= Dose,N= #voxels) and DTs (xn= DT,N= #dwell positions), mean differences (MD) in organD2ccand high-risk CTV D90 over all patients (where positive indicates higher clinical dose), and mean Dice similarity coefficients (DSC) for 100% isodose contours.Main results. Validation plans agreed well with clinical plans (MADdose= 1.1%, MADDT= 4 s or 0.8% of total plan time,D2ccMD = -0.2% to 0.2% and D90 MD = -0.6%, DSC = 0.99). For automated plans, MADdose= 6.5% and MADDT= 10.3 s (2.1%). The slightly higher clinical metrics in automated plans (D2ccMD = -3.8% to 1.3% and D90 MD = -5.1%) were due to higher neural network dose predictions. The overall shape of the automated dose distributions were similar to clinical doses (DSC = 0.91).Significance. Automated planning with 3D dose predictions could provide significant time savings and standardize treatment planning across practitioners, regardless of experience.

Advances in Automated Treatment Planning

Treatment planning in radiation therapy has progressed enormously over the past several decades. Such advancements came in the form of innovative hardware and algorithms, giving rise to modalities such as intensity-modulated radiation therapy and volume modulated arc therapy, greatly improving patient outcome and quality of life. While these developments have improved the overall plan quality, they have also given rise to higher treatment planning complexity. This has resulted in increased treatment planning time and higher variability in the final approved plan quality. Radiation oncology, as an already technologically advanced field, has much research and implementation involving the use of AI. The field has begun to show the efficacy of using such technologies in many of its sub-areas, such as in diagnosis, imaging, segmentation, treatment planning, quality assurance, treatment delivery, and follow-up. Some AI technologies have already been clinically implemented by commercial systems. In this article, we will provide an overview to methods involved with treatment planning in radiation therapy. In particular, we will review the recent research and literature related to automation of the treatment planning process, leading to potentially higher efficiency and higher quality plans. We will then present the current and future challenges, as well as some future perspectives.

Domain knowledge driven 3D dose prediction using moment-based loss function

Objective. To propose a novel moment-based loss function for predicting 3D dose distribution for the challenging conventional lung intensity modulated radiation therapy plans. The moment-based loss function is convex and differentiable and can easily incorporate clinical dose volume histogram (DVH) domain knowledge in any deep learning (DL) framework without computational overhead. Approach. We used a large dataset of 360 (240 for training, 50 for validation and 70 for testing) conventional lung patients with 2 Gy × 30 fractions to train the DL model using clinically treated plans at our institution. We trained a UNet like convolutional neural network architecture using computed tomography, planning target volume and organ-at-risk contours as input to infer corresponding voxel-wise 3D dose distribution. We evaluated three different loss functions: (1) the popular mean absolute error (MAE) loss, (2) the recently developed MAE + DVH loss, and (3) the proposed MAE + moments loss. The quality of the predictions was compared using different DVH metrics as well as dose-score and DVH-score, recently introduced by the AAPM knowledge-based planning grand challenge. Main results. Model with (MAE + moment) loss function outperformed the model with MAE loss by significantly improving the DVH-score (11%, p < 0.01) while having similar computational cost. It also outperformed the model trained with (MAE + DVH) by significantly improving the computational cost (48%) and the DVH-score (8%, p < 0.01). Significance. DVH metrics are widely accepted evaluation criteria in the clinic. However, incorporating them into the 3D dose prediction model is challenging due to their non-convexity and non-differentiability. Moments provide a mathematically rigorous and computationally efficient way to incorporate DVH information in any DL architecture. The code, pretrained models, docker container, and Google Colab project along with a sample dataset are available on our DoseRTX GitHub (https://github.com/nadeemlab/DoseRTX)

Artificial intelligence in radiotherapy

Radiotherapy is a discipline closely integrated with computer science. Artificial intelligence (AI) has developed rapidly over the past few years. With the explosive growth of medical big data, AI promises to revolutionize the field of radiotherapy through highly automated workflow, enhanced quality assurance, improved regional balances of expert experiences, and individualized treatment guided by multi-omics. In addition to independent researchers, the increasing number of large databases, biobanks, and open challenges significantly facilitated AI studies on radiation oncology. This article reviews the latest research, clinical applications, and challenges of AI in each part of radiotherapy including image processing, contouring, planning, quality assurance, motion management, and outcome prediction. By summarizing cutting-edge findings and challenges, we aim to inspire researchers to explore more future possibilities and accelerate the arrival of AI radiotherapy.

A Survey on Deep Learning for Precision Oncology

Precision oncology, which ensures optimized cancer treatment tailored to the unique biology of a patient’s disease, has rapidly developed and is of great clinical importance. Deep learning has become the main method for precision oncology. This paper summarizes the recent deep-learning approaches relevant to precision oncology and reviews over 150 articles within the last six years. First, we survey the deep-learning approaches categorized by various precision oncology tasks, including the estimation of dose distribution for treatment planning, survival analysis and risk estimation after treatment, prediction of treatment response, and patient selection for treatment planning. Secondly, we provide an overview of the studies per anatomical area, including the brain, bladder, breast, bone, cervix, esophagus, gastric, head and neck, kidneys, liver, lung, pancreas, pelvis, prostate, and rectum. Finally, we highlight the challenges and discuss potential solutions for future research directions.

Combining dense elements with attention mechanisms for 3D radiotherapy dose prediction on head and neck cancers

Purpose

External radiation therapy planning is a highly complex and tedious process as it involves treating large target volumes, prescribing several levels of doses, as well as avoiding irradiating critical structures such as organs at risk close to the tumor target. This requires highly trained dosimetrists and physicists to generate a personalized plan and adapt it as treatment evolves, thus affecting the overall tumor control and patient outcomes. Our aim is to achieve accurate dose predictions for head and neck (H&N) cancer patients on a challenging in‐house dataset that reflects realistic variability and to further compare and validate the method on a public dataset.

Methods

We propose a three‐dimensional (3D) deep neural network that combines a hierarchically dense architecture with an attention U‐net (HDA U‐net). We investigate a domain knowledge objective, incorporating a weighted mean squared error (MSE) with a dose‐volume histogram (DVH) loss function. The proposed HDA U‐net using the MSE‐DVH loss function is compared with two state‐of‐the‐art U‐net variants on two radiotherapy datasets of H&N cases. These include reference dose plans, computed tomography (CT) information, organs at risk (OARs), and planning target volume (PTV) delineations. All models were evaluated using coverage, homogeneity, and conformity metrics as well as mean dose error and DVH curves.

Results

Overall, the proposed architecture outperformed the comparative state‐of‐the‐art methods, reaching 0.95 (0.98) on D95 coverage, 1.06 (1.07) on the maximum dose value, 0.10 (0.08) on homogeneity, 0.53 (0.79) on conformity index, and attaining the lowest mean dose error on PTVs of 1.7% (1.4%) for the in‐house (public) dataset. The improvements are statistically significant (p<0.05) for the homogeneity and maximum dose value compared with the closest baseline. All models offer a near real‐time prediction, measured between 0.43 and 0.88 s per volume.

Conclusion

The proposed method achieved similar performance on both realistic in‐house data and public data compared to the attention U‐net with a DVH loss, and outperformed other methods such as HD U‐net and HDA U‐net with standard MSE losses. The use of the DVH objective for training showed consistent improvements to the baselines on most metrics, supporting its added benefit in H&N cancer cases. The quick prediction time of the proposed method allows for real‐time applications, providing physicians a method to generate an objective end goal for the dosimetrist to use as reference for planning. This could considerably reduce the number of iterations between the two expert physicians thus reducing the overall treatment planning time.

Towards a safe and efficient clinical implementation of machine learning in radiation oncology by exploring model interpretability, explainability and data-model dependency

The interest in machine learning (ML) has grown tremendously in recent years, partly due to the performance leap that occurred with new techniques of deep learning, convolutional neural networks for images, increased computational power, and wider availability of large datasets. Most fields of medicine follow that popular trend and, notably, radiation oncology is one of those that are at the forefront, with already a long tradition in using digital images and fully computerized workflows. ML models are driven by data, and in contrast with many statistical or physical models, they can be very large and complex, with countless generic parameters. This inevitably raises two questions, namely, the tight dependence between the models and the datasets that feed them, and the interpretability of the models, which scales with its complexity. Any problems in the data used to train the model will be later reflected in their performance. This, together with the low interpretability of ML models, makes their implementation into the clinical workflow particularly difficult. Building tools for risk assessment and quality assurance of ML models must involve then two main points: interpretability and data-model dependency. After a joint introduction of both radiation oncology and ML, this paper reviews the main risks and current solutions when applying the latter to workflows in the former. Risks associated with data and models, as well as their interaction, are detailed. Next, the core concepts of interpretability, explainability, and data-model dependency are formally defined and illustrated with examples. Afterwards, a broad discussion goes through key applications of ML in workflows of radiation oncology as well as vendors' perspectives for the clinical implementation of ML.

Modeling physician's preference in treatment plan approval of stereotactic body radiation therapy of prostate cancer

Objective.Treatment planning of radiation therapy is a time-consuming task. It is desirable to develop automatic planning approaches to generate plans favorable to physicians. The purpose of this study is to develop a deep learning based virtual physician network (VPN) that models physician's preference on plan approval for prostate cancer stereotactic body radiation therapy (SBRT).Approach.VPN takes one planning target volume (PTV) and eight organs at risk structure images, as well as a dose distribution of a plan seeking approval as input. It outputs a probability of approving the plan, and a dose distribution indicating improvements to the input dose. Due to the lack of unapproved plans in our database, VPN is trained using an adversarial framework. 68 prostate cancer patients who received 45Gyin 5-fraction SBRT were selected in this study, with 60 patients for training and cross validation, and 8 patients for independent testing.Main results.The trained VPN was able to differentiate approved and unapproved plans with Area under the curve 0.97 for testing data. For unapproved plans, after applying VPN's suggested dose improvement, the improved dose agreed with ground truth with relative differences2.03±2.17%for PTVD98%,0.49±0.29%for PTVV95%,3.08±2.24%for penile bulbDmean,3.73±2.20%for rectumV50%,and2.06±1.73%for bladderV50%.Significance.VPN was developed to accurately model a physician's preference on plan approval and to provide suggestions on how to improve the dose distribution.

Toward automatic beam angle selection for pencil-beam scanning proton liver Treatments: A deep learning-based approach

BACKGROUND

Dose deposition characteristics of proton radiation can be advantageous over photons. Proton treatment planning however poses additional challenges for the planners. Proton therapy is usually delivered with only a small number of beam angles, and the quality of a proton treatment plan is largely determined by the beam angles employed. Finding the optimal beam angles for a proton treatment plan requires time and experience, motivating the investigation of automatic beam angle selection methods.

PURPOSE

A deep learning-based approach to automatic beam angle selection is proposed for proton pencil-beam scanning treatment planning of liver lesions.

METHODS

We cast beam-angle selection as a multi-label classification problem. To account for angular boundary discontinuity, the underlying convolution neural network is trained with the proposed Circular Earth Mover's Distance based regularization and multi-label circular-smooth label technique. Furthermore, an analytical algorithm emulating proton treatment planners' clinical practice is employed in post-processing to improve the output of the model. Forty-nine patients that received proton liver treatments between 2017 and 2020 were randomly divided into training (n = 31), validation (n = 7), and test sets (n = 11). AI-selected beam angles were compared with those angles selected by human planners, and the dosimetric outcome was investigated by creating plans using knowledge-based treatment planning.

RESULTS

For 7 of the 11 cases in the test set, AI-selected beam angles agreed with those chosen by human planners to within 20 degrees (median angle difference = 10°; mean = 18.6°). Moreover, out of the total 22 beam angles predicted by the model, 15 (68%) were within 10 degrees of the human-selected angles. The high correlation in beam angles resulted in comparable dosimetric statistics between proton treatment plans generated using AI- and human-selected angles. For the cases with beam angle differences exceeding 20°, the dosimetric analysis showed similar plan quality although with different emphases on organ-at-risk sparing.

CONCLUSIONS

This pilot study demonstrated the feasibility of a novel deep learning-based beam angle selection technique. Testing on liver cancer patients showed that the resulting plans were clinically viable with comparable dosimetric quality to those using human-selected beam angles. In tandem with auto-contouring and knowledge-based treatment planning tools, the proposed model could represent a pathway for nearly fully automated treatment planning in proton therapy. This article is protected by copyright. All rights reserved.

Robust automated radiation therapy treatment planning using scenario-specific dose prediction and robust dose mimicking

Purpose

We present a framework for robust automated treatment planning using machine learning, comprising scenario‐specific dose prediction and robust dose mimicking.

Methods

The scenario dose prediction pipeline is divided into the prediction of nominal dose from input image and the prediction of scenario dose from nominal dose, each using a deep learning model with U‐net architecture. By using a specially developed dose–volume histogram–based loss function, the predicted scenario doses are ensured sufficient target coverage despite the possibility of the training data being non‐robust. Deliverable plans may then be created by solving a robust dose mimicking problem with the predictions as scenario‐specific reference doses.

Results

Numerical experiments are performed using a data set of 52 intensity‐modulated proton therapy plans for prostate patients. We show that the predicted scenario doses resemble their respective ground truth well, in particular while having target coverage comparable to that of the nominal scenario. The deliverable plans produced by the subsequent robust dose mimicking were showed to be robust against the same scenario set considered for prediction.

Conclusions

We demonstrate the feasibility and merits of the proposed methodology for incorporating robustness into automated treatment planning algorithms.

Probabilistic Pareto plan generation for semiautomated multicriteria radiation therapy treatment planning

Objective.We propose a semiautomatic pipeline for radiation therapy treatment planning, combining ideas from machine learning-automated planning and multicriteria optimization (MCO).Approach.Using knowledge extracted from historically delivered plans, prediction models for spatial dose and dose statistics are trained and furthermore systematically modified to simulate changes in tradeoff priorities, creating a set of differently biased predictions. Based on the predictions, an MCO problem is subsequently constructed using previously developed dose mimicking functions, designed in such a way that its Pareto surface spans the range of clinically acceptable yet realistically achievable plans as exactly as possible. The result is an algorithm outputting a set of Pareto optimal plans, either fluence-based or machine parameter-based, which the user can navigate between in real time to make adjustments before a final deliverable plan is created.Main results.Numerical experiments performed on a dataset of prostate cancer patients show that one may often navigate to a better plan than one produced by a single-plan-output algorithm.Significance.We demonstrate the potential of merging MCO and a data-driven workflow to automate labor-intensive parts of the treatment planning process while maintaining a certain extent of manual control for the user.

Artificial Intelligence in Radiation Therapy

Artificial intelligence (AI) has great potential to transform the clinical workflow of radiotherapy. Since the introduction of deep neural networks, many AI-based methods have been proposed to address challenges in different aspects of radiotherapy. Commercial vendors have started to release AI-based tools that can be readily integrated to the established clinical workflow. To show the recent progress in AI-aided radiotherapy, we have reviewed AI-based studies in five major aspects of radiotherapy including image reconstruction, image registration, image segmentation, image synthesis, and automatic treatment planning. In each section, we summarized and categorized the recently published methods, followed by a discussion of the challenges, concerns, and future development. Given the rapid development of AI-aided radiotherapy, the efficiency and effectiveness of radiotherapy in the future could be substantially improved through intelligent automation of various aspects of radiotherapy.

Assessment of efficacy in automated plan generation for Varian Ethos intelligent optimization engine

Varian Ethos, a new treatment platform, is capable of automatically generating treatment plans for initial planning and for online, adaptive planning, using an intelligent optimization engine (IOE). The primary purpose of this study is to assess the efficacy of Varian Ethos IOE for auto‐planning and intercompare different treatment modalities within the Ethos treatment planning system (TPS). A total of 36 retrospective prostate and proximal seminal vesicles cases were selected for this study. The prescription dose was 50.4 Gy in 28 fractions to the proximal seminal vesicles, with a simultaneous integrated boost of 70 Gy to the prostate gland. Based on RT intent, three treatment plans were auto‐generated in the Ethos TPS and were exported to the Eclipse TPS for intercomparison with the Eclipse treatment plan. When normalized for the same planning target volume (PTV) coverage, Ethos plans D_max% were 108.1 ± 1.2%, 108.4 ± 1.6%, and 109.6 ± 2.0%, for the 9‐field IMRT, 12‐field IMRT, and 2‐full arc VMAT plans, respectively. This compared well with Eclipse plan D_max% values, which was 108.8 ± 1.4%. OAR indices were also evaluated for Ethos plans using Radiation Therapy Oncology Group report 0415 as a guide and were found to be comparable to each other and the Eclipse plans. While all Ethos plans were comparable, we found that, in general, the Ethos 12‐field IMRT plans met most of the dosimetric goals for treatment. Also, Ethos IOE consistently generated dosimetrically hotter VMAT plans versus IMRT plans. On average, Ethos TPS took 13 min to generate 2‐full arc VMAT plans, compared to 5 min for 12‐field IMRT plans. Varian Ethos TPS can generate multiple treatment plans in an efficient time frame and the quality of the plans could be deemed clinically acceptable when compared to manually generated treatment plans.

A hybrid optimization strategy for deliverable Intensity-modulated radiotherapy plan generation using deep learning-based dose prediction

PURPOSE

To propose a clinically feasible automatic planning solution for external beam intensity-modulated radiotherapy (IMRT) including dose prediction via a deep learning and voxel-based optimization strategy.

MATERIALS AND METHODS

The dose distribution of patients was predicted using a U-Net-based deep learning network based on the patient's anatomy information. One hundred seventeen patients with nasopharyngeal cancer (NPC) and 200 patients with rectal cancer were enrolled in this study. For NPC cases, 94 cases were included in the training dataset, 13 in the validation dataset and 10 in the testing dataset. For rectal cancer cases, 172 cases were included in the training set, 18 in the validation set and 10 in the testing set. A voxel-based optimization strategy, "Voxel", was proposed to achieve treatment planning optimization by dividing body voxels into two parts: inside planning target volumes (PTVs) and outside PTVs. Fixed dose-volume objectives were attached to the total objective function to realize individualized planning intended as the "hybrid" optimizing strategy. Automatically generated plans were compared with clinically approved plans to evaluate clinical gains, according to dosimetric indices and dose-volume histograms (DVHs).

RESULTS

Similarities were found between the DVH of the predicted dose and clinical plan, although significant differences were found in some organs at risk (OARs). Better organ sparing and suboptimal PTV coverage were shown using the voxel strategy; however, the deviation in homogeneity indices (HI) and conformity indices (CI) of the PTV between automatically generated plans and manual plans were reduced by the hybrid strategy ((Manual plans)/(Voxel plans)/(Hybrid plans): HI of PTV70 (1.06/1.12/1.02), CI of PTV70 (0.79/0.58/0.76)). The optimization time for each patient was within 1 minute and included fluence map optimization, leaf sequencing and control point optimization. All the generated plans (voxel & hybrid strategy) could be delivered on uRT-linac 506c (United Imaging Health care, Shanghai, China).

CONCLUSION

Deliverable plans can be generated by incorporating a voxel-based optimization strategy into a commercial TPS. The hybrid optimization method shows the benefit and clinical feasibility in generating clinically acceptable plans. This article is protected by copyright. All rights reserved.

Radiomics outperforms clinical factors in characterizing human papilloma virus (HPV) for patients with oropharyngeal squamous cell carcinomas

Purpose:To utilize radiomic features extracted from CT images to characterize Human Papilloma Virus (HPV) for patients with oropharyngeal cancer squamous cell carcinoma (OPSCC).Methods:One hundred twenty-eight OPSCC patients with known HPV-status (60-HPV+ and 68-HPV-, confirmed by immunohistochemistry-P16-protein testing) were retrospectively studied. Radiomic features (11 feature-categories) were extracted in 3D from contrast-enhanced (CE)-CT images of gross-tumor-volumes using 'in-house' software ('ROdiomiX') developed and validated following the image-biomarker-standardization-initiative (IBSI) guidelines. Six clinical factors were investigated: Age-at-Diagnosis, Gender, Total-Charlson, Alcohol-Use, Smoking-History, and T-Stage. A Least-Absolute-Shrinkage-and-Selection-Operation (Lasso) technique combined with a Generalized-Linear-Model (Lasso-GLM) were applied to perform regularization in the radiomic and clinical feature spaces to identify the ranking of optimal feature subsets with most representative information for prediction of HPV. Lasso-GLM models/classifiers based on clinical factors only, radiomics only, and combined clinical and radiomics (ensemble/integrated) were constructed using random-permutation-sampling. Tests of significance (One-way ANOVA), average Area-Under-Receiver-Operating-Characteristic (AUC), and Positive and Negative Predictive values (PPV and NPV) were computed to estimate the generalization-error and prediction performance of the classifiers.Results:Five clinical factors, including T-stage, smoking status, and age, and 14 radiomic features, including tumor morphology, and intensity contrast were found to be statistically significant discriminators between HPV positive and negative cohorts. Performances for prediction of HPV for the 3 classifiers were: Radiomics-Lasso-GLM: AUC/PPV/NPV=0.789/0.755/0.805; Clinical-Lasso-GLM: 0.676/0.747/0.672, and Integrated/Ensemble-Lasso-GLM: 0.895/0.874/0.844. Results imply that the radiomics-based classifier enabled better characterization and performance prediction of HPV relative to clinical factors, and that the combination of both radiomics and clinical factors yields even higher accuracy characterization and predictive performance.Conclusion:Albeit subject to confirmation in a larger cohort, this pilot study presents encouraging results in support of the role of radiomic features towards characterization of HPV in patients with OPSCC.

Dose-based radiomic analysis (dosiomics) for intensity-modulated radiotherapy in patients with prostate cancer: Correlation between planned dose distribution and biochemical failure

PURPOSE

Although radiotherapy is one of the most significant modalities for localized prostate cancer, the prognostic factors for biochemical recurrence (BCR) regarding the treatment plan are unclear. We aimed to develop a novel dosiomics-based prediction model for BCR in patients with prostate cancer and clarify the correlations between the dosimetric factors and BCR.

METHODS AND MATERIALS

This study included 489 patients with localized prostate cancer (BCR: 96, No-BCR: 393) who received intensity-modulated radiation therapy. A total of 2,475 dosiomic features were extracted from the dose distributions on the prostate, clinical target volume (CTV), and planning target volume. A prediction model for BCR was trained on a training cohort of 342 patients. The performance of this model was validated using the concordance index (C-index) in a validation cohort of 147 patients. Another model was constructed using clinical variables, dosimetric parameters, and radiomic features for comparisons. Kaplan-Meier curves with log-rank analysis were used to assess the univariate discrimination based on the predictive dosiomic features.

RESULTS

The dosiomic feature derived from the CTV was significantly associated with BCR (hazard ratio: 0.73; 95% confidence interval [CI]: 0.57-0.93; P = .01). Although the dosiomics model outperformed the dosimetric and radiomics models, it did not outperform the clinical model. The performance significantly improved by combining the clinical variables and dosiomic features (C-index: 0.67; 95% CI: 0.65-0.68; P < .0001). The predictive dosiomic features were used to distinguish high-risk and low-risk patients (P < .05).

CONCLUSIONS

The dosiomic feature extracted from the CTV was significantly correlated with BCR in patients with prostate cancer, and the dosiomics model outperformed the model with conventional dose indices. Hence, new metrics for evaluating the quality of a treatment plan are warranted. Moreover, further research should be conducted to determine whether dosiomics can be incorporated in a clinical workflow or clinical trial.

Knowledge-based radiation treatment planning: A data-driven method survey

This paper surveys the data-driven dose prediction methods investigated for knowledge-based planning (KBP) in the last decade. These methods were classified into two major categories-traditional KBP methods and deep-learning (DL) methods-according to their techniques of utilizing previous knowledge. Traditional KBP methods include studies that require geometric or anatomical features to either find the best-matched case(s) from a repository of prior treatment plans or to build dose prediction models. DL methods include studies that train neural networks to make dose predictions. A comprehensive review of each category is presented, highlighting key features, methods, and their advancements over the years. We separated the cited works according to the framework and cancer site in each category. Finally, we briefly discuss the performance of both traditional KBP methods and DL methods, then discuss future trends of both data-driven KBP methods to dose prediction.

Probabilistic feature extraction, dose statistic prediction and dose mimicking for automated radiation therapy treatment planning

PURPOSE

We propose a general framework for quantifying predictive uncertainties of dose-related quantities and leveraging this information in a dose mimicking problem in the context of automated radiation therapy treatment planning.

METHODS

A three-step pipeline, comprising feature extraction, dose statistic prediction and dose mimicking, is employed. In particular, the features are produced by a convolutional variational autoencoder and used as inputs in a previously developed nonparametric Bayesian statistical method, estimating the multivariate predictive distribution of a collection of predefined dose statistics. Specially developed objective functions are then used to construct a probabilistic dose mimicking problem based on the produced distributions, creating deliverable treatment plans.

RESULTS

The numerical experiments are performed using a dataset of 94 retrospective treatment plans of prostate cancer patients. We show that the features extracted by the variational autoencoder capture geometric information of substantial relevance to the dose statistic prediction problem and are related to dose statistics in a more regularized fashion than hand-crafted features. The estimated predictive distributions are reasonable and outperforms a non-input-dependent benchmark method, and the deliverable plans produced by the probabilistic dose mimicking agree better with their clinical counterparts than for a non-probabilistic formulation.

CONCLUSIONS

We demonstrate that prediction of dose-related quantities may be extended to include uncertainty estimation and that such probabilistic information may be leveraged in a dose mimicking problem. The treatment plans produced by the proposed pipeline resemble their original counterparts well, illustrating the merits of a holistic approach to automated planning based on probabilistic modeling.

A feasibility study on deep learning-based individualized 3D dose distribution prediction

PURPOSE

Radiation therapy treatment planning is a trial-and-error, often time-consuming process. An approximately optimal dose distribution corresponding to a specific patient's anatomy can be predicted by using pre-trained deep learning (DL) models. However, dose distributions are often optimized based not only on patient-specific anatomy but also on physicians' preferred trade-offs between planning target volume (PTV) coverage and organ at risk (OAR) sparing or among different OARs. Therefore, it is desirable to allow physicians to fine-tune the dose distribution predicted based on patient anatomy. In this work, we developed a DL model to predict the individualized 3D dose distributions by using not only the patient's anatomy but also the desired PTV/OAR trade-offs, as represented by a dose volume histogram (DVH), as inputs.

METHODS

In this work, we developed a modified U-Net network to predict the 3D dose distribution by using patient PTV/OAR masks and the desired DVH as inputs. The desired DVH, fine-tuned by physicians from the initially predicted DVH, is first projected onto the Pareto surface, then converted into a vector, and then concatenated with feature maps encoded from the PTV/OAR masks. The network output for training is the dose distribution corresponding to the Pareto optimal DVH. The training/validation datasets contain 77 prostate cancer patients, and the testing dataset has 20 patients.

RESULTS

The trained model can predict a 3D dose distribution that is approximately Pareto optimal while having the DVH closest to the input desired DVH. We calculated the difference between the predicted dose distribution and the optimized dose distribution that has a DVH closest to the desired one for the PTV and for all OARs as a quantitative evaluation. The largest absolute error in mean dose was about 3.6% of the prescription dose, and the largest absolute error in the maximum dose was about 2.0% of the prescription dose.

CONCLUSIONS

In this feasibility study, we have developed a 3D U-Net model with the patient's anatomy and the desired DVH curves as inputs to predict an individualized 3D dose distribution that is approximately Pareto optimal while having the DVH closest to the desired one. The predicted dose distributions can be used as references for dosimetrists and physicians to rapidly develop a clinically acceptable treatment plan.

A hierarchical deep reinforcement learning framework for intelligent automatic treatment planning of prostate cancer intensity modulated radiation therapy

Purpose.We have previously proposed an intelligent automatic treatment planning (IATP) framework that builds a virtual treatment planner network (VTPN) to operate a treatment planning system (TPS) to generate high-quality radiation therapy (RT) treatment plans. While the potential of IATP in automating RT treatment planning has been demonstrated, its poor scalability caused by an almost linear growth of network size with the number of treatment planning parameters (TPPs) is a bottleneck, preventing its application in complicate, but clinically relevant treatment planning problems. The decision-making behavior of the trained network is hard to understand. Motivated by the decision-making process of a human planner, this study proposes a hierarchical IATP framework.Methods and materials.The hierarchical VTPN (HieVTPN) consists of three networks, i.e. Structure-Net, Parameter-Net, and Action-Net. When interacting with a TPS, the networks are employed in a sequential order in each step to decide the structure to adjust, the TPP to adjust for the selected structure, and the specific adjustment manner for the parameter, respectively. We developed an end-to-end hierarchical deep reinforcement learning scheme to simultaneously train the three networks. We then evaluated the effectiveness of the proposed framework in the treatment planning problems for prostate cancer intensity modulated RT (IMRT) and stereotactic body RT (SBRT). We benchmarked the performance of our approach by comparing plans made by VTPN of a parallel architecture, and the human plans submitted for competition in the 2016 American Association of Medical Dosimetrist (AAMD)/Radiosurgery Society (RSS) Plan Study. We analyzed scalability of the network size with respect to the number of TPPs. Numerical experiments were also performed to understand the rationale of the decision-making behaviors of the trained HieVTPN.Results.Both HieVTPNs for prostate IMRT and SBRT were trained successfully using 10 training patient cases and 5 validation cases. For IMRT, HieVTPN was able to generate high-quality plans for 59 testing patient cases that were not included in training process, achieving an average plan score of 8.62 (±0.83), with 9 being the maximal score. The score was comparable to that of the VTPN, 8.45 (±0.48). For SBRT planning, HieVTPN achieved an average plan score of 139.07 on five testing patient cases compared to the score of 132.21 averaged over the human plans summited for competition in AAMD/RSS plan study. Different from VTPN with network size linearly scaling with the number of TPPs, the network size of HieVTPN is almost independent of the number of TPPs. It was also observed that the decision-making behaviors of HieVTPN were understandable and generally agreed with the human experience.Conclusions.With the scalability and explainability, the hierarchical IATP framework is more favorable than the previous framework in terms of handling treatment planning problems involving a large number of TPPs.

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

Purpose

To present the technical details of the runner‐up model in the open knowledge‐based planning (OpenKBP) challenge for the dose–volume histogram (DVH) stream. The model was designed to ensure simple and reproducible training, without the necessity of costly advanced generative adversarial network (GAN) techniques.

Methods

The model was developed based on the OpenKBP challenge dataset, consisting of 200 and 40 head‐and‐neck patients for training and validation, respectively. The final model is a U‐Net with additional ResNet blocks between up‐ and down convolutions. The results were obtained by training the model with AdamW with the One Cycle scheduler. The loss function is a combination of the L1 loss with a feature loss, which uses a pretrained video classifier as a feature extractor. The performance was evaluated on another 100 patients in the OpenKBP test dataset. The DVH metrics of the test data were evaluated, where D0.1cc, and Dmean were calculated for the organs at risk (OARs) and D1%, D95%, and D99% were computed for the target structures. DVH metric differences between predicted and true dose are reported in percentage.

Results

The model achieved 2nd and 4th place in the DVH and dose stream of the OpenKBP challenge, respectively. The dose and DVH score were 2.62 ± 1.10 and 1.52 ± 1.06, respectively. Mean dose differences for the different structures and DVH parameters were within ±1%.

Conclusion

This straightforward approach produced excellent results. It incorporated One Cycle Learning, ResNet, and feature‐based losses, which are common computer vision techniques.

Learning-based dose prediction for pancreatic stereotactic body radiation therapy using dual pyramid adversarial network

Treatment planning for pancreatic cancer stereotactic body radiation therapy (SBRT) is very challenging owing to vast spatial variations and close proximity of many organs-at-risk. Recently, deep learning (DL) based methods have been applied in dose prediction tasks of various treatment sites with the aim of relieving planning challenges. However, its effectiveness on pancreatic cancer SBRT is yet to be fully explored due to limited investigations in the literature. This study aims to further current knowledge in DL-based dose prediction tasks by implementing and demonstrating the feasibility of a new dual pyramid networks (DPNs) integrated DL model for predicting dose distributions of pancreatic SBRT. The proposed framework is composed of four parts: CT-only feature pyramid network (FPN), contour-only FPN, late fusion network and an adversarial network. During each phase of the network, combination of mean absolute error, gradient difference error, histogram matching, and adversarial loss is used for supervision. The performance of proposed model was demonstrated for pancreatic cancer SBRT plans with doses prescribed between 33 and 50 Gy across as many as three planning target volumes (PTVs) in five fractions. Five-fold cross validation was performed on 30 patients, and another 20 patients were used as holdout tests of trained model. Predicted plans were compared with clinically approved plans through dose volume parameters and two-paired t-test. For the same sets, our results were compared with three different DL architectures: 3D U-Net, 3D U-Net with adversarial learning, and DPN without adversarial learning. The proposed framework was able to predict 87% and 91% of clinically relevant dose parameters for cross validation sets and holdout sets, respectively, without any significant differences (P > 0.05). Dose distribution predicted by our framework was also able to predict the intentional hotspots as feature characteristics of SBRT plans. Our method achieved higher correlation coefficients with the ground truth in 22/26, 24/26 and 20/26 dose volume parameters compared to the network without adversarial learning, 3D U-Net, and 3D U-Net with adversarial learning, respectively. Overall, the proposed model was able to predict doses to cases with both single and multiple PTVs. In conclusion, the DPN integrated DL model was successfully implemented, and demonstrated good dose prediction accuracy and dosimetric characteristics for pancreatic cancer SBRT.

Artificial intelligence and machine learning for medical imaging: A technology review

Artificial intelligence (AI) has recently become a very popular buzzword, as a consequence of disruptive technical advances and impressive experimental results, notably in the field of image analysis and processing. In medicine, specialties where images are central, like radiology, pathology or oncology, have seized the opportunity and considerable efforts in research and development have been deployed to transfer the potential of AI to clinical applications. With AI becoming a more mainstream tool for typical medical imaging analysis tasks, such as diagnosis, segmentation, or classification, the key for a safe and efficient use of clinical AI applications relies, in part, on informed practitioners. The aim of this review is to present the basic technological pillars of AI, together with the state-of-the-art machine learning methods and their application to medical imaging. In addition, we discuss the new trends and future research directions. This will help the reader to understand how AI methods are now becoming an ubiquitous tool in any medical image analysis workflow and pave the way for the clinical implementation of AI-based solutions.

A comparison of Monte Carlo dropout and bootstrap aggregation on the performance and uncertainty estimation in radiation therapy dose prediction with deep learning neural networks

Recently, artificial intelligence technologies and algorithms have become a major focus for advancements in treatment planning for radiation therapy. As these are starting to become incorporated into the clinical workflow, a major concern from clinicians is not whether the model is accurate, but whether the model can express to a human operator when it does not know if its answer is correct. We propose to use Monte Carlo Dropout (MCDO) and the bootstrap aggregation (bagging) technique on deep learning (DL) models to produce uncertainty estimations for radiation therapy dose prediction. We show that both models are capable of generating a reasonable uncertainty map, and, with our proposed scaling technique, creating interpretable uncertainties and bounds on the prediction and any relevant metrics. Performance-wise, bagging provides statistically significant reduced loss value and errors in most of the metrics investigated in this study. The addition of bagging was able to further reduce errors by another 0.34% for [Formula: see text] and 0.19% for [Formula: see text] on average, when compared to the baseline model. Overall, the bagging framework provided significantly lower mean absolute error (MAE) of 2.62, as opposed to the baseline model's MAE of 2.87. The usefulness of bagging, from solely a performance standpoint, does highly depend on the problem and the acceptable predictive error, and its high upfront computational cost during training should be factored in to deciding whether it is advantageous to use it. In terms of deployment with uncertainty estimations turned on, both methods offer the same performance time of about 12 s. As an ensemble-based metaheuristic, bagging can be used with existing machine learning architectures to improve stability and performance, and MCDO can be applied to any DL models that have dropout as part of their architecture.

Improving efficiency of training a virtual treatment planner network via knowledge-guided deep reinforcement learning for intelligent automatic treatment planning of radiotherapy

PURPOSE

We previously proposed an intelligent automatic treatment planning framework for radiotherapy, in which a virtual treatment planner network (VTPN) is built using deep reinforcement learning (DRL) to operate a treatment planning system (TPS) by adjusting treatment planning parameters in it to generate high-quality plans. We demonstrated the potential feasibility of this idea in prostate cancer intensity-modulated radiation therapy (IMRT). Despite the success, the process to train a VTPN via the standard DRL approach with an ϵ-greedy algorithm was time-consuming. The required training time was expected to grow with the complexity of the treatment planning problem, preventing the development of VTPN for more complicated but clinically relevant scenarios. In this study, we proposed a novel knowledge-guided DRL (KgDRL) approach that incorporated knowledge from human planners to guide the training process to improve the efficiency of training a VTPN.

METHOD

Using prostate cancer IMRT as a test bed, we first summarized a number of rules in the actions of adjusting treatment planning parameters of our in-house TPS. During the training process of VTPN, in addition to randomly navigating the large state-action space, as in the standard DRL approach using the ϵ-greedy algorithm, we also sampled actions defined by the rules. The priority of sampling actions from rules decreased over the training process to encourage VTPN to explore new policy on parameter adjustment that were not covered by the rules. To test this idea, we trained a VTPN using KgDRL and compared its performance with another VTPN trained using the standard DRL approach. Both networks were trained using 10 training patient cases and five additional cases for validation, while another 59 cases were employed for the evaluation purpose.

RESULTS

It was found that both VTPNs trained via KgDRL and standard DRL spontaneously learned how to operate the in-house TPS to generate high-quality plans, achieving plan quality scores of 8.82 (±0.29) and 8.43 (±0.48), respectively. Both VTPNs outperformed treatment planning purely based on the rules, which had a plan score of 7.81 (±1.59). VTPN trained with eight episodes using KgDRL was able to perform similar to that trained using DRL with 100 epochs. The training time was reduced from more than a week to ~13 hrs.

CONCLUSION

The proposed KgDRL framework was effective in accelerating the training process of a VTPN by incorporating human knowledge, which will facilitate the development of VTPN for more complicated treatment planning scenarios.

Artificial intelligence-based radiotherapy machine parameter optimization using reinforcement learning

PURPOSE

To develop and evaluate a volumetric modulated arc therapy (VMAT) machine parameter optimization (MPO) approach based on deep-Q reinforcement learning (RL) capable of finding an optimal machine control policy using previous prostate cancer patient CT scans and contours, and applying the policy to new cases to rapidly produce deliverable VMAT plans in a simplified beam model.

METHODS

A convolutional deep-Q network was employed to control the dose rate and multileaf collimator of a C-arm linear accelerator model using the current dose distribution and machine parameter state as input. A Q-value was defined as the discounted cumulative cost based on dose objectives, and experience-replay RL was performed to determine a policy to minimize the Q-value. A two-dimensional network design was employed which optimized each opposing leaf pair independently while monitoring the corresponding dose plane blocked by those leaves. This RL approach was applied to CT and contours from 40 retrospective prostate cancer patients. The dataset was split into training (15 patients) and validation (5 patients) groups to optimize the network, and its performance was tested in an independent cohort of 20 patients by comparing RL-based dose distributions to conformal arcs and clinical intensity modulated radiotherapy (IMRT) delivering a prescription dose of 78 Gy in 40 fractions.

RESULTS

Mean ± SD execution time of the RL VMAT optimization was 1.5 ± 0.2 s per slice. In the test cohort, mean ± SD (P-value) planning target volume (PTV), bladder, and rectum dose were 80.5 ± 2.0 Gy (P < 0.001), 44.2 ± 14.6 Gy (P < 0.001), and 43.7 ± 11.1 Gy (P < 0.001) for RL VMAT compared to 81.6 ± 1.1 Gy, 51.6 ± 12.9 Gy, and 36.0 ± 12.3 Gy for clinical IMRT.

CONCLUSIONS

RL was applied to VMAT MPO using clinical patient contours without independently optimized treatment plans for training and achieved comparable target and normal tissue dose to clinical plans despite the application of a relatively simple network design originally developed for video-game control. These results suggest that extending a RL approach to a full three-dimensional beam model could enable rapid artificial intelligence-based optimization of deliverable treatment plans, reducing the time required for radiotherapy planning without requiring previous plans for training.

Technical note: Interpolated Pareto surface similarity metrics for multi-criteria optimization in radiation therapy

PURPOSE

There is a strong clinical need to evaluate different multi-criteria optimization (MCO) algorithms, including inverse optimization sampling algorithms and machine learning-based predictions. This study aims to develop and compare several interpolated Pareto surface similarity metrics.

MATERIALS AND METHODS

The first metric is the root-mean-square error (RMSE) evaluated between vertices on the interpolated surfaces, augmented by intra-simplex sampling of the barycentric coordinates of the surfaces' simplicial complexes. The second metric is the average projected distance (APD), which evaluates the displacements between the vertices and computes their projections along the mean displacement. The third metric is the average nearest-point distance (ANPD), which numerically integrates point-to-simplex distances over the sampled simplices of the interpolated surfaces. These metrics were compared by their convergence rates, the times required to achieve convergence, and their representation of the underlying surface interpolations. For analysis, several interpolated Pareto surface pairs were constructed abstractly, with one pair from a nasopharyngeal treatment planning case using MCO.

RESULTS

Convergence within 1% is typically achieved at approximately 50 and 80 samples per barycentric dimension for the RMSE and the ANPD, respectively. Calculation requires approximately 1 and 10 ms to achieve convergence for the RMSE and the ANPD in two dimensions, respectively, while the APD always requires < 1 ms. These time costs are much higher in higher dimensions for just the RMSE and ANPD. The APD values more closely approximated the ANPD limits than the RMSE limits.

CONCLUSION

The ANPD's formulation and generality make it likely more meaningful than the RMSE and APD for representing the similarity between the underlying interpolated surfaces rather than the sampling points on the surfaces. However, in situations requiring high-speed evaluations, the APD may be more desirable due to its speed, independence from a subjectively chosen sampling rate, and similarity to the ANPD limits.

Using deep learning to predict beam-tunable Pareto optimal dose distribution for intensity-modulated radiation therapy

PURPOSE

Many researchers have developed deep learning models for predicting clinical dose distributions and Pareto optimal dose distributions. Models for predicting Pareto optimal dose distributions have generated optimal plans in real time using anatomical structures and static beam orientations. However, Pareto optimal dose prediction for intensity-modulated radiation therapy (IMRT) prostate planning with variable beam numbers and orientations has not yet been investigated. We propose to develop a deep learning model that can predict Pareto optimal dose distributions by using any given set of beam angles, along with patient anatomy, as input to train the deep neural networks. We implement and compare two deep learning networks that predict with two different beam configuration modalities.

METHODS

We generated Pareto optimal plans for 70 patients with prostate cancer. We used fluence map optimization to generate 500 IMRT plans that sampled the Pareto surface for each patient, for a total of 35 000 plans. We studied and compared two different models, Models I and II. Although they both used the same anatomical structures - including the planning target volume (PTV), organs at risk (OARs), and body - these models were designed with two different methods for representing beam angles. Model I directly uses beam angles as a second input to the network as a binary vector. Model II converts the beam angles into beam doses that are conformal to the PTV. We divided the 70 patients into 54 training, 6 validation, and 10 testing patients, thus yielding 27 000 training, 3000 validation, and 5000 testing plans. Mean square loss (MSE) was taken as the loss function. We used the Adam optimizer with a default learning rate of 0.01 to optimize the network's performance. We evaluated the models' performance by comparing their predicted dose distributions with the ground truth (Pareto optimal) dose distribution, in terms of dose volume histogram (DVH) plots and evaluation metrics such as PTV D98 , D95 , D50 , D2 , Dmax , Dmean , Paddick Conformation Number, R50, and Homogeneity index.

RESULTS

Our deep learning models predicted voxel-level dose distributions that precisely matched the ground truth dose distributions. The DVHs generated also precisely matched the ground truth. Evaluation metrics such as PTV statistics, dose conformity, dose spillage (R50), and homogeneity index also confirmed the accuracy of PTV curves on the DVH. Quantitatively, Model I's prediction error of 0.043 (confirmation), 0.043 (homogeneity), 0.327 (R50), 2.80% (D95), 3.90% (D98), 0.6% (D50), and 1.10% (D2) was lower than that of Model II, which obtained 0.076 (confirmation), 0.058 (homogeneity), 0.626 (R50), 7.10% (D95), 6.50% (D98), 8.40% (D50), and 6.30% (D2). Model I also outperformed Model II in terms of the mean dose error and the max dose error on the PTV, bladder, rectum, left femoral head, and right femoral head.

CONCLUSIONS

Treatment planners who use our models will be able to use deep learning to control the trade-offs between the PTV and OAR weights, as well as the beam number and configurations in real time. Our dose prediction methods provide a stepping stone to building automatic IMRT treatment planning.

Fully automated dose prediction using generative adversarial networks in prostate cancer patients

Purpose

Although dose prediction for intensity modulated radiation therapy (IMRT) has been accomplished by a deep learning approach, delineation of some structures is needed for the prediction. We sought to develop a fully automated dose-generation framework for IMRT of prostate cancer by entering the patient CT datasets without the contour information into a generative adversarial network (GAN) and to compare its prediction performance to a conventional prediction model trained from patient contours.

Methods

We propose a synthetic approach to translate patient CT datasets into a dose distribution for IMRT. The framework requires only paired-images, i.e., patient CT images and corresponding RT-doses. The model was trained from 81 IMRT plans of prostate cancer patients, and then produced the dose distribution for 9 test cases. To compare its prediction performance to that of another trained model, we created a model trained from structure images. Dosimetric parameters for the planning target volume (PTV) and organs at risk (OARs) were calculated from the generated and original dose distributions, and mean differences of dosimetric parameters were compared between the CT-based model and the structure-based model.

Results

The mean differences of all dosimetric parameters except for D_98% and D_95% for PTV were within approximately 2% and 3% of the prescription dose for OARs in the CT-based model, while the differences in the structure-based model were within approximately 1% for PTV and approximately 2% for OARs, with a mean prediction time of 5 seconds per patient.

Conclusions

Accurate and rapid dose prediction was achieved by the learning of patient CT datasets by a GAN-based framework. The CT-based dose prediction could reduce the time required for both the iterative optimization process and the structure contouring, allowing physicians and dosimetrists to focus their expertise on more challenging cases.