Machine learning techniques for medical image analysis with data scarcity

Abstract

Medical imaging is integral to modern healthcare for diagnosing, treating, and monitoring various conditions. There is considerable interest in leveraging machine learning to derive insights from medical imaging, particularly in tasks such as image classification, segmentation, and anomaly detection. Unlike natural image recognition, medical image analysis faces challenges due to the limited annotated data availability and the complexity of integrating multi-modal medical information.

This thesis aims to develop machine/deep learning techniques to address these challenges, improving diagnostic efficiency and reducing the labour-intensive nature of current medical practices. We designed a data-based few-shot learning scheme to investigate the use of pre-trained deep learning models to extract meaningful data representations, focusing on scenarios where data are sparse relative to feature dimensions. A novel approach using non-negative matrix factorization (NMF), particularly discriminative variants like DNMF and SCNMFS, is explored for dimensionality reduction in low-data settings typical of medical inference tasks.

Additionally, we propose a method for integrating multi-modal medical data to generate standardized medical reports. The proposed 'data-text-data' transformation strategy enhances interpretability and accuracy by converting input indicators into sequential word-embedded representations and then reconstructing them into their original format, ensuring clinically relevant outcomes.

Moreover, to address the scarcity of pixel-level annotations in medical imaging, we introduce a diffusion model with discrepancy-based features. This approach translates inconsistencies in image-level annotations into distribution discrepancies among heterogeneous samples while preserving information within homogeneous samples. Unlike traditional segmentation methods that rely heavily on pairwise annotations, this method enhances segmentation accuracy by implicitly leveraging annotation distributions and generative learning paradigms within medical data.

Overall, these contributions aim to advance the application of machine/deep learning in medical imaging, addressing challenges related to data scarcity and the complexity of integrating multi-modal medical information in clinical settings.

More information

Identifiers

ORCID for Keqiang Fan: orcid.org/0000-0002-9411-2892

ORCID for Xiaohao Cai: orcid.org/0000-0003-0924-2834

ORCID for Mahesan Niranjan: orcid.org/0000-0001-7021-140X

Catalogue record

Export record