Organ at risk segmentation in head and neck CT scans

Abstract

Head and neck cancer is one of the most common cancers worldwide responsible for around half a million deaths each year. One of the most common types of treatment for head and neck cancer includes radiation therapy which uses high doses of radiation to target the tumour mass and high risk areas, inadvertently affecting nearby healthy cells. These nearby healthy cells are know as organs at risk and locating them precisely is detrimental to minimise side effects upon them. In the past few decades, automatic segmentation has risen in popularity especially in the medical field where structures in the body can be contoured automatically by a model. More recently deep neural networks have been used to automatically segment these Organs at risk using different implementations. By reviewing current state of the art technologies, three approaches were developed to assess and compare how well the different methods segment a comprehensive set of OARs. These included a 2D U-Net, a 3D U-Net and a multi-layer 3D U-Net. The 2D U-Net is a simple 2D encoder decoder network. The 3D U-Net expands on this by adding an extra dimension to the U-Net allowing volumetric data for training. The multi-layer 3D U-Net uses multiple models to first segment easier OARs so they can act as landmarks to segment simpler OARs. The 2D model, 3D model, and multi-layer 3D model achieved an average dice score of 75.7%, 80.24% and 83.8% respectively on 15 OARs.