Back to search

University of Southampton

Inference from medical images with linear discriminant analysis and deep learning-derived features

Abstract

dc:description.abstract

Several challenges arise in the application of modern computer vision and machine learning techniques to making inferences from medical images. They include relatively low sample images, biases inherent in the training data that can act as confounding variables, and the need to incorporate clinical variables about patients. In this dissertation, we combine deep neural networks, pre-trained on large datasets of natural images (e.g. ImageNet) acting as feature extractors, and linear discriminant subspaces to address these issues. We start with our extension to the classic linear discriminant analysis (LDA) to derive multiple mutually orthogonal discriminant directions in the multi-class discriminant subspace, which we refer to as generalised optimal LDA (GO-LDA). Unlike previous work in the field of LDA, the discriminability of the subspace we derive is not limited by the number of classes in the multi-class problem. Empirical work on 14 datasets covering 11 different disease domains shows the advantage of this discriminant subspace approach in a few-shot learning setting. Furthermore, bias in data, arising from confounding information from protected characteristics (e.g. gender or skin tone color), which can lead to unacceptable decision barriers in society, is of serious concern in medical inference problems. Here, we show how an approach biased towards deriving orthogonal discriminant directions, whereby one direction separates the protected characteristic and one separates disease state, can effectively address this issue. This is demonstrated using dermatology and chest X-ray problems in which skin tone color and gender induce confounding issues. Finally, we demonstrate that appending relevant clinical variables in the reduced discriminant subspace is effective, as demonstrated using tuberculosis and dermatology datasets.

Degree

thesis:*
Name dc:type.qualificationname
Ph.D.
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
University of Southampton
Year dc:date.issued
2024

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Liu, Jiahui
Advisors dc:contributor.advisor
  • Cai, Xiaohao
  • Niranjan, Mahesan

Rights

Language dc:language
English, eng

Chain of custody

source
Harvested from
University of Southampton
Base URL
eprints.soton.ac.uk/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Liu, Jiahui. Inference from medical images with linear discriminant analysis and deep learning-derived features. doctoral thesis, University of Southampton, 2024.