Deep Learning Approaches for the Assessment of Glaucoma Progression Using Fundus Photographs and Visual Field Data

Reduanul Haque

doi:10.60867/00000160

Glaucoma is characterized as a progressive optic neuropathy and a leading cause of irreversible vision loss worldwide. Assessing progression, i.e., accurate detection and prediction of future progression, is crucial for timely intervention and preserving vision. However, traditional clinical approaches, including sequential visual field (VF) assessments and fundus photographic (FP) examinations, are often limited by variability and sparse longitudinal data. In addition, these approaches rely on manual grading by experts, which is a subjective process. Moreover, the data distribution is inherently imbalanced due to a lower number of progression cases. To address these challenges, this thesis presents several deep learning (DL) networks, each targeting a specific aspect of detecting and predicting future glaucoma progression using sequential FPs and VF data. As such, the thesis has the following contributions: First, we propose a two-stage computational learning pipeline to estimate sector-wise rate of changes in VF sensitivity values, which were subsequently used to detect progression only from sequential FPs, establishing a foundation for automated progression assessment. Additionally, a novel dataset construction method is introduced that aligns temporal fundus image sequences with vector-valued labels derived from visual field regression analysis. Second, we use stereoscopic sequential FPs to detect progression using depth-aware, cross-interaction Siamese network. The proposed network integrates depth-aware, patch diffusion module with cross interaction network that is sensitive to subtle, localized structural changes while maintaining a low number of trainable parameters to prevent overfitting on a small and imbalanced dataset. Third, we propose an anatomically guided, fluctuation-aware bidirectional long shortterm memory (Bi-LSTM) network to predict future VFs and progression from prior sequential VF data. The method incorporates anatomically coherent clustering of VF sensitivity values using a structure–function mapping between VF test locations and optic nerve head sectors. This design substantially improves prediction accuracy, particularly for localized and highly fluctuating glaucomatous damage. Finally, we employ a CNN-LSTM network with class-weighted focal loss to address the inherent class imbalance in glaucoma progression datasets and to predict future VF progression using prior sequential FPs. Moreover, a temporal sliding window technique is used to construct the dataset, and several labelled datasets were constructed using three, four, and five sequential fundus photographs. For progression labels, either two or three VF test locations having a statistically significant regression slope are used. Our results demonstrate the efficacy of DL approaches that can effectively learn complex spatiotemporal relationships from sequential ophthalmic data. This thesis highlights the potential of AI-driven, DL-based sequential modelling for clinical decision support in glaucoma management and establishes a framework for future research on sequential data in ophthalmology.

Deep Learning Approaches for the Assessment of Glaucoma Progression Using Fundus Photographs and Visual Field Data

Files and links (1)

Abstract

Details

Metrics