Towards real-world clinical colonoscopy deep learning models for video-based bowel preparation and generalisable polyp segmentation

Haithami, Mahmood (2024) Towards real-world clinical colonoscopy deep learning models for video-based bowel preparation and generalisable polyp segmentation. PhD thesis, University of Nottingham.

[thumbnail of Haithami, Mahmood, 20219960, final version.pdf]
Preview
PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (6MB) | Preview

Abstract

Colorectal cancer is the most prevalence type of cancers within the digestive system. Early screening and removal of precancerous growths in the colon decrease mortality rate. The golden standard screening type for colon is colonoscopy which is conducted by a medical expert (i.e., colonoscopist). Nevertheless, due to human biases, fatigue, and experience level of the colonoscopist, colorectal cancer missing rate is negatively affected. Artificial intelligence (AI) methods hold immense promise not just in automating colonoscopy tasks but also enhancing the performance of colonoscopy screening in general. The recent development of intense computational GPUs enabled a computational-demanding AI method (i.e., deep learning) to be utilised in various medical applications. However, given the gap between the clinical-practice and the proposed deep learning models in the literature, the actual effectiveness of such methods is questionable. Hence, this thesis highlights such gaps that arises from the separation between the theoretical and practical aspect of deep learning methods applied to colonoscopy. The aim is to evaluate the current state of deep learning models applied in colonoscopy from a clinical angle, and accordingly propose better evaluation strategies and deep learning models. The aim is translated into three distinct objectives. The first objective is to develop a systematic evaluation method to assess deep learning models from a clinical perspective. The second objective is to develop a novel deep learning architecture that leverages spatial information within colonoscopy videos to enhance the effectiveness of deep learning models on real-clinical environments. The third objective is to enhance the generalisability of deep learning models on unseen test images by developing a novel deep learning framework. To translate these objectives into practice, two critical colonoscopy tasks, namely, automatic bowel preparation and polyp segmentation are attacked. In both tasks, subtle overestimations are found in the literature and discussed in the thesis theoretically and demonstrated empirically. These overestimations are induced by improper validation sets that would not appear or represent the real-world clinical environment. Arbitrary dividing colonoscopy datasets to do deep learning evaluation can result in producing similar distributions, hence, achieving unrealistic results. Accordingly, these factors are considered in the thesis to avoid such subtle overestimation. For the automatic bowel preparation task, colonoscopy videos that closely resemble clinical settings are considered as input and accordingly it necessitates the design of the proposed model as well as evaluation experiments. The proposed model’s architecture is designed to utilise both temporal and spatial information within colonoscopy videos using Gated Recurrent Unit (GRU) and a proposed Multiplexer unit, respectively. Meanwhile for the polyp segmentation task, the efficiency of current deep learning models is tested in terms of their generalisation capabilities using unseen test sets from different medical centres. The proposed framework consists of two connected models. The first model is responsible for gradually transforming textures of input images and arbitrary change their colours. Meanwhile the second model is a segmentation model that outlines polyp regions. Exposing the segmentation model to such transformed images acquires the segmentation model texture/colour invariant properties, hence, enhances the generalisability of the segmentation model. In this thesis, rigorous experiments are conducted to evaluate the proposed models against the state-of-the-art models. The yielded results indicate that the proposed models outperformed the state-of-the-art models under different settings.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Yi, Iman Liao
Ahmed, Amr
Amin, Hafeez Ullah
Keywords: deep learning, colonoscopy, artificial intelligence, digestive system, gastrointestinal tract, polyps segmentation, bowel preparations, classification, videos, images
Subjects: Q Science > Q Science (General)
Faculties/Schools: University of Nottingham, Malaysia > Faculty of Science and Engineering — Science > School of Computer Science
Item ID: 78094
Depositing User: Haithami, Mahmood
Date Deposited: 27 Jul 2024 04:40
Last Modified: 27 Jul 2024 04:40
URI: https://eprints.nottingham.ac.uk/id/eprint/78094

Actions (Archive Staff Only)

Edit View Edit View