Multi-Layer Anatomically-Guided Diffusion and Adaptive Calibration for Precision Radiotherapy Dose Prediction
Abstract
Purpose
Dose distribution maps are critical for high-quality radiotherapy treatment planning, yet plan optimization remains largely manual and trial-and-error, leading to substantial time costs and limited efficiency. While knowledge-based and deep-learning-based dose prediction methods have been proposed to assist medical physicists, many do not explicitly account for inter-slice anatomical information that influences dose deposition, and their clinical utility is constrained by inadequate dose accuracy within critical regions of interest (ROIs). This study proposes a diffusion-based framework integrating multi-layer anatomical guidance and adaptive ROI-focused calibration to improve global and ROI-specific dose prediction accuracy.
Methods
We propose MAGDiff-ADC, a conditional diffusion framework that formulates dose prediction as a generative denoising process: starting from random noise, the model progressively refines a dose map conditioned on computed tomography (CT) images and ROI masks. A conditional U-Net is equipped with a multi-layer spatially-adaptive normalization mechanism to hierarchically fuse anatomical features from adjacent slices, explicitly modeling complex 3D dose–anatomy correlations. We also introduce a lightweight, fully convolutional Adaptive Dose Calibration (ADC) module to refine dose values within ROIs and mitigate pixel-level inaccuracies of diffusion models. MAGDiff-ADC is evaluated on a public benchmark and an in-house dataset using multiple metrics, emphasizing ROI-level performance.
Results
Across both datasets, MAGDiff-ADC achieves state-of-the-art performance against prior methods. Multi-layer anatomical guidance improves inter-slice dependency modeling, producing more anatomically consistent dose distributions. The ADC module further enhances dose accuracy in clinically critical ROIs, outperforming previous approaches across multiple metrics.
Conclusion
MAGDiff-ADC offers an effective diffusion-based solution for radiotherapy dose prediction by combining adjacent-slice anatomical guidance with ROI-targeted adaptive calibration. It improves dose accuracy, especially within critical ROIs, enhancing the clinical utility of dose prediction to support efficient, high-quality treatment planning.