Hybrid 3D-ResNet Deep Learning Model for Automatic Segmentation of Thoracic Organs at Risk in CT Images

In image radiation therapy, accurate segmentation of organs at risk (OARs) is a very essential task and has clinical applications in cancer treatment. The segmentation of organs close to lung, breast, or esophageal cancer is a routine and time-consuming process. The automatic segmentation of organs at risk would be an essential part of treatment planning for patients suffering radiotherapy. The position and shape variation, morphology inherent and low soft tissue contrast between neighboring organs across each patient’s scans is the challenging task for automatic segmentation of OARs in Computed Tomography (CT) images. The objective of this paper is to use automatic segmentation of the organs near risk in CT images using deep learning model. The paper proposes a hybrid 3D-ResNet based deep learning model with Atrous spatial pyramid pooling module and Project & Excite (PE)' module for 3D volumetric segmentation using Thoracic Organs at Risk (SegTHOR) dataset. The proposed model produces better results as compared to state-of-the-art deep learning models used in SegTHOR dataset. Proposed 3D volumetric Hybrid deep model could be used for automatic segmentation of OARs in clinical applications and would be helpful to diagnose lung, breast or esophageal cancer in CT images.