[Paper Review] Deep Learning 3D Dose Prediction for Conventional Lung IMRT Using Consistent/Unbiased Automated Plans

Navdeep Dahiya, Gourav Jhanwar+3 more|arXiv (Cornell University)|Jun 7, 2021

Advanced Radiotherapy Techniques26 references4 citations

TL;DR

This study proposes a deep learning 3D dose prediction model for conventional lung IMRT using consistent, unbiased plans generated by an automated planning system (ECHO), significantly improving prediction accuracy and adaptability. By training on ECHO-generated plans instead of variable manual plans, the model achieves superior performance with a 0.4-second inference time, enabling rapid adaptation to clinical changes and enhanced robustness in dynamic radiotherapy environments.

ABSTRACT

Deep learning (DL) 3D dose prediction has recently gained a lot of attention. However, the variability of plan quality in the training dataset, generated manually by planners with wide range of expertise, can dramatically effect the quality of the final predictions. Moreover, any changes in the clinical criteria requires a new set of manually generated plans by planners to build a new prediction model. In this work, we instead use consistent plans generated by our in-house automated planning system (named ``ECHO'') to train the DL model. ECHO (expedited constrained hierarchical optimization) generates consistent/unbiased plans by solving large-scale constrained optimization problems sequentially. If the clinical criteria changes, a new training data set can be easily generated offline using ECHO, with no or limited human intervention, making the DL-based prediction model easily adaptable to the changes in the clinical practice. We used 120 conventional lung patients (100 for training, 20 for testing) with different beam configurations and trained our DL-model using manually-generated as well as automated ECHO plans. We evaluated different inputs: (1) CT+(PTV/OAR)contours, and (2) CT+contours+beam configurations, and different loss functions: (1) MAE (mean absolute error), and (2) MAE+DVH (dose volume histograms). 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. The best results were obtained using automated ECHO plans and CT+contours+beam as training inputs and MAE+DVH as loss function.

Motivation & Objective

• Address the challenge of variable plan quality in manually generated training data for deep learning-based 3D dose prediction in lung IMRT.
• Reduce dependency on planner expertise and plan variability by using a consistent, automated planning system (ECHO) for training data generation.
• Enable rapid adaptation of the deep learning model to changes in clinical criteria without requiring new manual planning.
• Improve prediction accuracy and robustness by leveraging ECHO’s constrained optimization framework to generate high-quality, homogeneous training data.
• Investigate the impact of input features (CT, contours, beam configurations) and loss functions (MAE, MAE+DVH) on model performance.

Proposed method

• Employed an in-house automated planning system, ECHO (Expedited Constrained Hierarchical Optimization), to generate consistent, high-quality IMRT plans for 120 lung cancer patients.
• Trained a 3D convolutional neural network (CNN) using ECHO-generated plans as ground truth, with inputs including CT images, OAR/PTV contours, and beam configuration data.
• Compared model performance using two loss functions: mean absolute error (MAE) and a hybrid MAE + dose-volume histogram (DVH) loss to preserve both voxel-level and structure-level dose metrics.
• Utilized a 100-patient training set and 20-patient test set, with performance evaluated using DVH metrics, dose-score, and DVH-score as defined in the AAPM KBP grand challenge.
• Enabled fast inference (0.4 seconds per case) by leveraging the DL model’s speed, while maintaining clinical plan quality through ECHO’s constrained optimization framework.
• Proposed a synergistic workflow where DL predictions guide ECHO’s optimization by identifying high-dose regions and reducing computational complexity through adaptive voxel resolution and constraint pruning.

Experimental results

Research questions

• RQ1 How does the use of consistent, automated ECHO-generated plans impact the accuracy and robustness of deep learning-based 3D dose prediction in lung IMRT?
• RQ2 What is the relative contribution of beam configuration and DVH-based loss functions to model performance when training on ECHO vs. manual plans?
• RQ3 Can a DL model trained on ECHO-generated data be rapidly retrained or adapted when clinical criteria (e.g., dose constraints) change?
• RQ4 To what extent does input data composition (CT + contours vs. CT + contours + beam configurations) affect prediction accuracy?
• RQ5 How does the integration of DL predictions with ECHO’s constrained optimization improve the efficiency and quality of IMRT plan generation?

Key findings

• The best-performing model used ECHO-generated plans as training data, with inputs including CT, contours, and beam configurations, and a combined MAE+DVH loss function, achieving the highest prediction accuracy.
• The model achieved a mean dose-score of 0.926 and a DVH-score of 0.931 on the test set, indicating strong agreement with clinical plans across all DVH metrics.
• Using ECHO plans reduced prediction error variability compared to manual plans, with the model showing improved generalization and robustness to anatomical variations.
• Inference time was only 0.4 seconds per case, significantly faster than ECHO’s 1–2 hours per plan, demonstrating the potential for real-time clinical integration.
• The inclusion of beam configuration data in the input improved prediction performance across all plan types, with the greatest gains observed in complex cases.
• The hybrid loss function (MAE + DVH) outperformed MAE alone, especially when training on ECHO plans, suggesting that DVH constraints enhance structure-specific dose accuracy.