Automated Segmentation of Head and Neck Cancer on Computed Tomography Imaging

Abstract

Head and neck cancer (HNC) is the sixth leading cause of cancer death worldwide. Diagnosis and treatment are aided by CT imaging, and tumour segmentation on CT is an important but time consuming part of treatment planning. In this thesis, we investigate different approaches for the segmentation of HNC on head and neck CT scans.

For a segmentation convolutional neural network (CNN), it is important to start with pretrained weights to achieve good performance. Most prior work uses weights obtained through classification pretraining tasks. We conjecture that a segmentation pretraining task may be better suited for medical image segmentation, as then the pretraining task is more closely related to the final task. We develop a novel self-supervised segmentation pretraining task which we then use to pretrain the model on unlabelled CT images prior to fine-tuning it on expert-labelled images.

We compare model performance after pretraining on this new task against existing pretraining methods, including out-of-domain pretraining using ImageNet and in-domain pretraining using the Jigsaw task. All in-domain pretraining, both using Jigsaw and the novel segmentation method created here, were performed on a composite pretraining dataset of over 618,000 CT images which was created by combining and preprocessing 8 separate medical imaging datasets.

We find that optimal performance is obtained with ImageNet out-of-domain pretraining, and this performance rivals previously published work which used PET-CT combination images rather than CT alone. Our novel pretraining segmentation task improves performance over random starting weights but does not exceed the performance of ImageNet pretraining. We conjecture that this might be because the pretraining dataset for our task is much smaller than the pretraining dataset used for ImageNet.