Deep Learning Enabled Framework for Explainable Colorectal Polyp Classification

Abstract

Cancer remains a leading global cause of death, with its burden increasing worldwide. Early detection and classification of polyps are crucial for reducing cancer risk and improving outcomes. Accurate classification of colorectal cancer (CRC) is vital for effective healthcare and diagnosis. Automated medical imaging aids in these diagnoses, but pathologists often face challenges like inexperience and visual fatigue. This research aims to develop a sensitive virtual biopsy tool to assist pathologists in overcoming issues related to the quality, diversity, and variability of biomedical data.

This thesis aims to develop and evaluate a novel framework for the efficient and accurate classification of CRC polyps from endoscopy images. It involves studying the existing techniques, designing and implementing the classification framework, and validating the model to enhance sensitivity, accuracy, and clinical applicability. This research utilises deep learning models, including CNNs, CGANs, Transformer-based models, XAI models, and Nonlinear Dimensionality Reduction. It incorporates transfer and ensemble learning, synthetic image generation, fine-grained feature extraction, model explainability, and manifold learning with regularisation to enhance classification performance. The approach also includes a comprehensive pipeline for data preprocessing, model training, and performance evaluation on real-world datasets.

This novel deep learning framework achieves state-of-the-art results in CRC polyp classification by generating high-quality synthetic images and extracting detailed features for an explainable model. It uses nonlinear dimensionality reduction to handle high dimensionality and enhance performance. The model demonstrates an average accuracy improvement of 18% and a 17% increase in sensitivity compared to traditional methods, with sensitivities of 0.95, 0.96, and 0.96 on the UCI, PICCOLO, and CPDC datasets, respectively. Its strong performance across various metrics suggests it is a promising tool for clinical integration, aiding gastroenterologists in polyp classification for polypectomy.

This thesis significantly advances CRC image classification by developing a framework that addresses multiple data challenges and improves classification performance. The model shows potential for clinical use as a virtual biopsy tool to aid pathologists and enhance patient outcomes. Future work will focus on integrating real-time data detection, handling inaccurate clinical data, and developing a semi-supervised learning model to train with limited labelled data.