Medical Image Classification using Deep Learning Techniques and Uncertainty Quantification

Senousy, Zakaria (2023) Medical Image Classification using Deep Learning Techniques and Uncertainty Quantification. Doctoral thesis, Birmingham City University.

Zakaria Senousy PhD Thesis published_Final version_Submitted Oct 2022_Final Award Mar 2023.pdf - Accepted Version

Download (2MB)


The emergence of medical image analysis using deep learning techniques has introduced multiple challenges in terms of developing robust and trustworthy systems for automated grading and diagnosis. Several works have been presented to improve classification performance. However, these methods lack the diversity of capturing different levels of contextual information among image regions, strategies to present diversity in learning by using ensemble-based techniques, or uncertainty measures for predictions generated from automated systems. Consequently, the presented methods provide sub-optimal results which is not enough for clinical practice. To enhance classification performance and introduce trustworthiness, deep learning techniques and uncertainty quantification methods are required to provide diversity in contextual learning and the initial stage of explainability, respectively.

This thesis aims to explore and develop novel deep learning techniques escorted by uncertainty quantification for developing actionable automated grading and diagnosis systems. More specifically, the thesis provides the following three main contributions. First, it introduces a novel entropy-based elastic ensemble of Deep Convolutional Neural Networks (DCNNs) architecture termed as 3E-Net for classifying grades of invasive breast carcinoma microscopic images. 3E-Net is based on a patch-wise network for feature extraction and image-wise networks for final image classification and uses an elastic ensemble based on Shannon Entropy as an uncertainty quantification method for measuring the level of randomness in image predictions. As the second contribution, the thesis presents a novel multi-level context and uncertainty-aware deep learning architecture named MCUa for the classification of breast cancer microscopic images. MCUa consists of multiple feature extractors and multi-level context-aware models in a dynamic ensemble fashion to learn the spatial dependencies among image patches and enhance the learning diversity. Also, the architecture uses Monte Carlo (MC) dropout for measuring the uncertainty of image predictions and deciding whether an input image is accurate based on the generated uncertainty score. The third contribution of the thesis introduces a novel model agnostic method (AUQantO) that establishes an actionable strategy for optimising uncertainty quantification for deep learning architectures. AUQantO method works on optimising a hyperparameter threshold, which is compared against uncertainty scores from Shannon entropy and MC-dropout. The optimal threshold is achieved based on single- and multi-objective functions which are optimised using multiple optimisation methods.

A comprehensive set of experiments have been conducted using multiple medical imaging datasets and multiple novel evaluation metrics to prove the effectiveness of our three contributions to clinical practice. First, 3E-Net versions achieved an accuracy of 96.15% and 99.50% on invasive breast carcinoma dataset. The second contribution, MCUa, achieved an accuracy of 98.11% on Breast cancer histology images dataset. Lastly, AUQantO showed significant improvements in performance of the state-of-the-art deep learning models with an average accuracy improvement of 1.76% and 2.02% on Breast cancer histology images dataset and an average accuracy improvement of 5.67% and 4.24% on Skin cancer dataset using two uncertainty quantification techniques. AUQantO demonstrated the ability to generate the optimal number of excluded images in a particular dataset.

Item Type: Thesis (Doctoral)
10 October 2022Submitted
6 March 2023Accepted
Uncontrolled Keywords: Medical Image Classification, Deep Learning, Uncertainty Quantification
Subjects: CAH10 - engineering and technology > CAH10-03 - materials and technology > CAH10-03-05 - biotechnology
CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science
Divisions: Doctoral Research College > Doctoral Theses Collection
Faculty of Computing, Engineering and the Built Environment > School of Computing and Digital Technology
Depositing User: Jaycie Carter
Date Deposited: 24 Mar 2023 16:43
Last Modified: 24 Mar 2023 16:43

Actions (login required)

View Item View Item


In this section...