Patch-Based 3D UNet for Head and Neck Tumor Segmentation with an Ensemble of Conventional and Dilated Convolutions

Abstract

Automatic segmentation of tumor eliminates problems associated with manual annotation of region-of-interest (ROI) from medical images, such as significant human efforts and inter-observer variability. Accurate segmentation of head and neck tumor has a tremendous potential for better radiation treatment planning for cancer (such as oropharyngeal cancer) and also for optimized patient care. In recent times, the development in deep learning models has been able to effectively and accurately perform segmentation tasks in semantic segmentation as well as in medical image segmentation. In medical imaging, different modalities focus on different properties and combining the information from them can improve the segmentation task. In this paper we developed a patch-based deep learning model to tackle the memory issue associated with training the network on 3D images. Furthermore, an ensemble of conventional and dilated convolutions was used to take advantage of both methods: the smaller receptive field of conventional convolution allows to capture finer details, whereas the larger receptive field of dilated convolution allows to capture better global information. Using patch-based 3D UNet with an ensemble of conventional and dilated convolution yield promising result, with a final dice score of 0.6911.