Subgroup-Specific Risk-Controlled Dose Estimation in Radiotherapy

TL;DR

This work addresses the need for fast, accurate voxel-wise dose estimation in MR-Linac radiotherapy with reliable uncertainty quantification. It combines the DeepDose neural network for dose prediction with a subgroup-aware risk-controlling framework (SG-RCPS) to produce prediction intervals that guarantee coverage not only globally but also across unknown subgroups, such as beam foreground and background regions. The authors derive the SG-RCPS algorithm, validate it on a real multi-organ RT dataset, and show that SG-RCPS achieves substantially better subgroup calibration than standard RCPS, particularly along the radiation beam where accurate uncertainty is crucial. The approach enhances safety and trustworthiness in high-stakes RT dosing and lays groundwork for extending risk guarantees to additional, underrepresented subgroups in clinical settings.

Abstract

Cancer remains a leading cause of death, highlighting the importance of effective radiotherapy (RT). Magnetic resonance-guided linear accelerators (MR-Linacs) enable imaging during RT, allowing for inter-fraction, and perhaps even intra-fraction, adjustments of treatment plans. However, achieving this requires fast and accurate dose calculations. While Monte Carlo simulations offer accuracy, they are computationally intensive. Deep learning frameworks show promise, yet lack uncertainty quantification crucial for high-risk applications like RT. Risk-controlling prediction sets (RCPS) offer model-agnostic uncertainty quantification with mathematical guarantees. However, we show that naive application of RCPS may lead to only certain subgroups such as the image background being risk-controlled. In this work, we extend RCPS to provide prediction intervals with coverage guarantees for multiple subgroups with unknown subgroup membership at test time. We evaluate our algorithm on real clinical planing volumes from five different anatomical regions and show that our novel subgroup RCPS (SG-RCPS) algorithm leads to prediction intervals that jointly control the risk for multiple subgroups. In particular, our method controls the risk of the crucial voxels along the radiation beam significantly better than conventional RCPS.

Figures (3)

• Figure 1: Overview. We use the DeepDose network kontaxis2020deepdose to convert a personalized RT plan defined by input CT scan, beam center distance, radiological depth, source distance map, and beam shape ($X_1$ to $X_5$) to a voxel-wise dose prediction $\hat{f}(x)$. Extending DeepDose by our novel subgroup risk-controlled prediction sets algorithm (SG-RCPS) allows to obtain a calibrated upper and lower bound for the dose ($\hat{u}(X)$ & $\hat{l}(X)$), as well as the voxel-wise size of the interval ($\hat{u}(X)-\hat{l}(X)$) which serves as final uncertainty measure.
• Figure 2: Tumor-specific risks for the original calibration method (left) and our method (right) for the total image (top row), the background radiation (middle row) and foreground radiation (bottom row).
• Figure 3: A representative example for a liver tumor visualizing the qualitative differences between the uncertainty intervals for the non-subgroup-specific calibration and our method. The uncertainty intervals provided by our method are significantly wider ($p < 0.001$) than the ones generated by classical RCPS. All values are given in Gray (Gy).