D3Net: A Distribution-Driven Deep Network for Radiotherapy Dose Prediction

Abstract

Radiotherapy is a primary treatment for cancers to apply sufficient radiation dose to the planning target volume (PTV) while minimizing dose hazards to the organs at risk (OARs). Recently, convolutional neural network (CNN) has automated radiotherapy plan making by directly predicting the dose distribution maps. However, existing CNN-based methods ignore two critical dose distribution characteristics, i.e., 1) the spatial distribution of different dose values and 2) dose differences in the interior and exterior PTV, resulting in suboptimal predictions. In this article, we propose a distribution-driven deep network, named D3Net, to achieve automatic dose prediction by simultaneously considering its spatial distribution and dose differences. Concretely, D3Net is constructed by a traditional CNN framework embedded with a transformer encoder to extract both local and global dosimetric information. To investigate the spatial distribution of different dose values, we present an innovative discrete multidose constraint to measure multiple dose values in the predicted dose map with discrete dose masks. Besides, we design a PTV-guided triplet constraint to utilize the explicit geometry of PTV to refine dose feature representations in the interior and exterior PTV, thus facilitating the dose differences. The proposed method is validated on the two clinical datasets, achieving $| {{\Delta }{D}}_{98} |$ values of 1.87 Gy for rectum (REC) cancer and 1.08 Gy for cervical cancer. The experimental results surpass those of other state-of-the-art (SOTA) methods, verifying that the predicted dose distribution of our method is more closed to the clinically approved one.