Deep Rectum Segmentation for Image Guided Radiation Therapy with Synthetic Data

Dimitrios Mallios, Xiaohao Cai

Abstract: Image guidance nowadays is a crucial component for doctors to facilitate the design of the planning radiation therapy dosage. The delineation of soft organs in the planning phase and during the radiation therapy is crucial for the treatment procedure. Deep Learning (DL) flourishes, presenting state-of-the-art results in challenging computer vision tasks; however, the lack of annotated data hardens the research advancements for medical applications. The research in this paper develops DL approaches for the segmentation of organs-at-risk and specifically from images retrieved from a computed tomography system during the radiation treatment of each patient. The proposed approaches are based on convolutional neural architectures trained with only a couple of thousand images, and can also be trained online, showing its learning ability from new patients. The lack of annotated data is also addressed with synthetic data generated by a modified GAN. Experimental results demonstrate the excellent performance of the proposed approaches in rectum segmentation task.