A deep learning approach to segmentation of nasopharyngeal carcinoma using computed tomography

Abstract

Automated segmentation of Nasopharyngeal carcinoma (NPC) plays a critical role in the radiotherapy or chemo-radiotherapy for this cancer. Despite their improved performance, most deep learning models designed for this segmentation task use either magnetic resonance imaging (MRI) or multimodality data as input. In this paper, we propose a deep learning based algorithm called NPC-Seg for the segmentation of NPC using computed tomography (CT), which is less expensive and more available than MRI. This algorithm uses the location-to-segmentation framework. In the location step, it fine-tunes the pre-trained ResNeXt-50 U-Net with a newly proposed recall preserved loss to roughly segment the gross tumor volume (GTV) of each NPC. In the segmentation step, it fine-tunes the ResNeXt-50 U-Net again, but using the Dice loss, to segment the bounding box region detected in the location step on a patch-by-patch basis. We have evaluated the proposed NPC-Seg algorithm on the StructSeg-NPC dataset. Our algorithm achieves the Dice similarity coefficient (DSC) of 62.88±8.12% on 50 training data in the ten-fold cross-validation, substantially outperforming three existing deep learning methods, and also achieves an average DSC of 61.81% on the testing dataset in the online validation.