Diffusion-Based Radiotherapy Dose Prediction Guided by Inter-Slice Aware Structure Encoding

Deep learning (DL) has successfully automated dose distribution prediction for radiotherapy planning, increasing both efficiency and quality. However, existing methods commonly utilize ${{\mathbf{L}}_1}$ or ${{\mathbf{L}}_2}$ loss to calculate the posterior average, thus heavily suffering from the over-smoothing problem. To address this, we propose a diffusion model-based method, named DiffDose, to automatically predict radiotherapy dose distribution for cancer patients. Specifically, our DiffDose model contains a forward process and a reverse process. In the forward process, DiffDose gradually adds small noise to dose distribution maps via multiple steps until converting them to pure Gaussian noise, and a noise predictor is simultaneously trained to estimate the noise added in each step. In the reverse process, DiffDose iteratively removes the noise from a pure Gaussian noise leveraging the well-trained noise predictor and finally outputs the predicted dose distribution maps. Concretely, to provide the model with essential structure information, we design a structure encoder to learn the anatomical knowledge from patients’ anatomy images, guiding the noise predictor to generate dose distribution maps that are aware of personalized structures. Considering the latent continuity and similarity among sliced anatomy images, an inter-slice interaction transformer (I2T) block is embedded in the structure encoder to capture such long-range dependency. Extensive experiments on an in-house dataset involving 130 rectum cancer cases validate the superiority of our method.