Polyp Segmentation in Colonoscopy Images with Convolutional Neural Networks

Guo, YunBo (2019) Polyp Segmentation in Colonoscopy Images with Convolutional Neural Networks. Doctoral thesis, University of Central Lancashire.

[thumbnail of Thesis Document]
PDF (Thesis Document) - Submitted Version
Available under License Creative Commons Attribution Non-commercial Share Alike.



The thesis looks at approaches to segmentation of polyps in colonoscopy images. The aim was to investigate and develop methods that are robust, accurate and computationally efficient and which can compete with the current state-of-the-art in polyp segmentation.
Colorectal cancer is one of the leading cause of cancer deaths worldwide. To decrease mortality, an assessment of polyp malignancy is performed during colonoscopy examination so polyps can be removed at an early stage. In current routine clinical practice, polyps are detected and delineated manually in colonoscopy images by highly trained clinicians. To automate these processes, machine learning and computer vision techniques have been utilised. They have been shown to improve polyp detectability and segmentation objectivity. However, polyp segmentation is a very challenging task due to inherent variability of polyp morphology and colonoscopy image appearance.
This research considers a range of approaches to polyp segmentation – seeking out those that offer a best compromise between accuracy and computational complexity. Based on analysis of existing machine learning and polyp image segmentation techniques, a novel hybrid deep learning segmentation method is proposed to alleviate the impact of the above stated challenges on polyp segmentation. The method consists of two fully convolutional networks. The first proposed network is based on a compact architecture with large receptive fields and multiple classification paths. The method performs well on most images, accurately segmenting polyps of diverse morphology and
appearance. However, this network is prone to misdetection of very small polyps. To solve this problem, a second network is proposed, which primarily aims to improve sensitivity to small polyp details by emphasising low-level image features.
In order to fully utilise information contained in the available training dataset, comprehensive data augmentation techniques are adopted. To further improve the performance of the proposed segmentation methods, test-time data augmentation is also implemented.
A comprehensive multi-criterion analysis of the proposed methods is provided. The result demonstrates that the new methodology has better accuracy and robustness than the current state-of-the-art, as proven by the outstanding performance at the 2017 and 2018 GIANA polyp segmentation challenges.

Repository Staff Only: item control page