Exploring Eye Image Data to Enhance the Etiological Diagnostic Algorithm for Uveitis

Uveitis encompasses a heterogeneous group of intraocular inflammatory diseases with more than 80 recognized etiologies. The clinical overlap between etiologies, combined with the rarity of individual entities, makes etiological diagnosis a highly expert-dependent task. Improving the accuracy and robustness of etiological diagnosis therefore remains an important clinical challenge. The host team has already developed a prototype etiological algorithm that integrates structured clinical variables. The algorithm currently treats ophthalmic imaging as a source of discrete, expert-extracted features rather than as raw, high-dimensional input. This limits the amount of diagnostic information that can be automatically extracted from ophthalmic images. This PhD project addresses that gap by directly incorporating image data into the modelling pipeline. Ocular imaging modalities, principally fundus photography, fluorescein angiography (FA), and optical coherence tomography (OCT), capture structural and pathological details that are pivotal in diagnosing uveitis and identifying its underlying causes. Retinal vasculitis patterns, choroidal infiltrates, macular edema, and disc changes are among the many image-level biomarkers that carry strong etiological signals. Unlocking these signals through advanced AI techniques constitutes the central objective of this PhD project. The first research objective will be to extract fundus, angiographic, and OCT images in DICOM format from the hospital information system, establishing a high-quality, curated imaging dataset for model training and evaluation. Then the second objective will be to develop an innovative semi-supervised image feature extraction pipeline that reduces dependency on costly manual annotations while maximizing the diagnostic information harvested from both labelled and unlabeled data. Segmentation models approaches will be investigated to identify clinically meaningful imaging features across the different imaging modalities. Semi-supervised learning strategies, including self-training, consistency regularization, and pseudo-label generation, will enable the models to effectively leverage large amounts of unlabeled images. Finally the final objective will be to integrate the extracted image-derived features into the multimodal diagnostic algorithm and quantitatively assess the improvement in etiological classification performance relative to the clinical-data-only baseline. The project will be carried out over three years, with specific objectives and milestones defined for each year. During the first year the focus will be on establishing the imaging dataset and developing a robust pipeline including preprocessing, segmentation, semi-supervised learning and feature extraction. During the second year the objective will be to validate and improve the developed approaches through clinical collaboration. During the third year, the work will focus on consolidation and clinical validation of the proposed multimodal diagnostic pipeline.