MacularAttentionNet: Retinal OCT Diagnostic Space

Deployment Notice: Unfortunately, Hugging Face no longer provides free dynamic Gradio container runtime options. As this was developed for an academic dissertation under a zero-cost infrastructure budget, the live interactive server could not be kept running continuously. The full model code, pre-trained weights, and metrics remain publicly available for reproduction.
GitHub Codebase Model Repository Card

Welcome to the public application tracking dashboard for MacularAttentionNet. This research project was engineered as part of an MSc Computer Science dissertation at Keele University to address optimization and explainability limits in computer vision-based retinal disease detection pipelines.

Research Abstract & Overview

Automated analysis of Optical Coherence Tomography (OCT) cross-sections has historically hit performance bottlenecks when relying on pure global spatial feature mapping. This dissertation demonstrates that integrating an engineered Convolutional Block Attention Module (CBAM) into a legacy ResNet50 baseline creates a localized spatial-channel focus that successfully removes diagnostic noise. The hybrid network optimizes feature extraction by calculating exactly what lesions matter before determining where they are structured on the retinal layer boundary grid.

Empirical Evaluation Matrix

The models were evaluated using identical validation metrics over the balanced 968-sample testing array extracted from the standard Kermany 2018 benchmark dataset. The performance summary below showcases the architectural progression:

Model Architecture Macro Precision Macro Recall Macro F1-Score Final Test Accuracy
ResNet50 Baseline 0.9969 0.9969 0.9969 99.69%
ConvNeXt-Tiny Baseline 0.9969 0.9969 0.9969 99.69%
MacularAttentionNet (ResNet50 + CBAM) 1.0000 1.0000 1.0000 100.00%

Key Discovery: Both vanilla ResNet50 and ConvNeXt-Tiny misclassified the exact same three samples due to subtle feature overlaps between early-stage micro-drusen deposits and micro-fluid tracks in CNV. The added dual-core attention layers within MacularAttentionNet effectively resolved these errors, yielding optimal convergence.

Clinical Diagnostic Scope

The systems are validated across four distinct ophthalmological profiles: