Tubular Structure-Aware Convolutional Neural Networks for Organ at Risks Segmentation in Cervical Cancer Radiotherapy

Cervical cancer is the most frequent cancer type among women worldwide and radiotherapy is the major clinical treatment. Organs in the radiation field are called Organ at Risks (OARs), which are prone to irreversible damage during radiotherapy. Therefore, accurate delineation of OARs is a critical step in ensuring radiotherapy dosimetry accuracy. However, currently existing deep learning-based cervical cancer OARs segmentation methods do not make full advantage of anatomical information. In this paper, we develop a novel tubular structure-aware deep convolutional network method integrating the tubular anatomical morphological features into a model for colon, small intestine and rectum in cervical cancer OARs. Firstly, a tubular filter based on variable annular Gaussian kernel and gradient detection was used to produce the tubular feature map. Secondary, tubular feature map concatenated with original image was input into the nnU-Net network for anatomical morphological information learning. Finally, we evaluated our proposed method on the clinical collection datasets with brachytherapy. Compared to the baseline model and state-of-the-art model, DSC and Recall were improved and the relative volume error (RVE) was reduced for the OARs with a tubular shape.