Multi-Task Learning for Integrated Automated Contouring and Voxel-Based Dose Prediction in Radiotherapy

TL;DR

This work tackles the problem of inefficiency in radiotherapy planning arising from treating automated contouring and dose prediction as separate tasks. It introduces Multi-Automated, a multi-task learning framework with a shared encoder and cross-task attention that jointly learns automated contouring and voxel-based dose prediction from CT images. Across prostate and head-and-neck (OpenKBP) datasets, Multi-Automated improves dose-prediction accuracy (DVH-MAE) and contouring metrics (Dice) relative to baseline and sequential models, with several DVH gains reaching statistical significance (e.g., Prostate $D$-metrics improved from $3.900$ Gy to $3.528$ Gy and OpenKBP from $11.650$ Gy to $10.109$ Gy). The approach reduces reliance on input contour quality and demonstrates robustness across sites, providing a promising path toward more efficient and accurate automated radiotherapy planning, though limitations include ROI scope and non-deliverable dose predictions. Future work aims to extend the framework to multiple treatment sites and integrate multi-model, multi-site capabilities for broader clinical deployment.

Abstract

Deep learning-based automated contouring and treatment planning has been proven to improve the efficiency and accuracy of radiotherapy. However, conventional radiotherapy treatment planning process has the automated contouring and treatment planning as separate tasks. Moreover in deep learning (DL), the contouring and dose prediction tasks for automated treatment planning are done independently. In this study, we applied the multi-task learning (MTL) approach in order to seamlessly integrate automated contouring and voxel-based dose prediction tasks, as MTL can leverage common information between the two tasks and be able able to increase the efficiency of the automated tasks. We developed our MTL framework using the two datasets: in-house prostate cancer dataset and the publicly available head and neck cancer dataset, OpenKBP. Compared to the sequential DL contouring and treatment planning tasks, our proposed method using MTL improved the mean absolute difference of dose volume histogram metrics of prostate and head and neck sites by 19.82% and 16.33%, respectively. Our MTL model for automated contouring and dose prediction tasks demonstrated enhanced dose prediction performance while maintaining or sometimes even improving the contouring accuracy. Compared to the baseline automated contouring model with the dice score coefficients of 0.818 for prostate and 0.674 for head and neck datasets, our MTL approach achieved average scores of 0.824 and 0.716 for these datasets, respectively. Our study highlights the potential of the proposed automated contouring and planning using MTL to support the development of efficient and accurate automated treatment planning for radiotherapy.

Figures (5)

• Figure 1: Architectures, inputs, and outputs of the (i) sequential human contouring and planning (Seq-Human), (ii) sequential automated contouring and planning (Seq-Automated), and proposed framework (iii) automated contouring and planning using multi-task learning (Multi-Automated). In Seq-Human and Seq-Automated, the input includes clinician labeled ground-truth (GT) and deep learning (DL) regions of interest (ROI), respectivley, both concatenated channel-wise with the CT image input. Multi-Automated uses only CT imaging as an input but includes an additional decoder for predicting ROI (output) simultaneously with dose distributions.
• Figure 2: This figure illustrates qualitative results of Prostate dataset, showing input CT scans, ground truth clinical dose distributions (Clinical Dose), and the dose distributions generated by sequential automated contouring and planning model (Seq-Automated) and the automated contouring and planning model using multi-task learning (Multi-Automated). The voxel-wise Mean Absolute Error (MAE) for both model outputs are provided. The figure presents best, median, and worst cases in terms of MAE from the test dataset
• Figure 3: In this figure, we present the results of automated DL contouring using Prostate dataset. The image showcases the ground truth labels (GT), and the DL contouring outputs from the baseline model (DL-Baseline) and the automated multi-task contouring and planning model (Multi-Automated). The figure includes representative examples, displaying best, median, and worst cases from our test dataset.
• Figure 4: This figure presents qualitative results from the OpenKBP dataset, focusing on the head and neck cancer, showing input CT scans, ground truth clinical dose distributions (Clinical Dose), and the dose distributions generated by sequential automated contouring and planning (Seq-Automated) and the automated contouring and planning model using multi-task learning (Multi-Automated). The Mean Absolute Error (MAE) for both model outputs are provided. The figure presents the best, median, and worst cases from the test dataset, starting from the top.
• Figure 5: In this figure, we present the results of DL automated contouring using OpenKBP dataset. The image showcases the ground truth labels (GT), and the automated contouring outputs from the baseline (DL-Baseline) and the automated contouring and planning model using multi-task learning (Multi-Automated). Additionally, we provide the Dice Score Coefficient for the segmentation output. The figure includes representative examples, displaying the best, median, and worst cases from our test dataset.