Introduction
Retinal disease diagnosis has traditionally relied on standard fundus photography, which captures a limited 45° field-of-view (FOV). While effective for central pathologies, this narrow window often misses critical indicators in the peripheral retina. Ultra-Wide-Field (UWF) imaging has emerged as a transformative technology, providing up to a 200° view in a single capture. However, the sheer scale and complexity of these images present significant hurdles for automated diagnostic systems.
The Challenge of UWF Imaging
UWF images cover a vast anatomical area, meaning that pathological features—such as tiny hemorrhages or microaneurysms—occupy a very small fraction of the total pixels. These lesions are often low-contrast and can be obscured by peripheral distortion, uneven illumination, or artifacts like eyelashes.
Traditional Convolutional Neural Networks (CNNs) typically struggle with this "needle in a haystack" problem. They often fail to balance the need for global structural analysis with the requirement for high-resolution local feature detection.
Proposed Methodology: A Dual-Path Approach
To address these challenges, researchers developed a Dual-Path and Multi-Scale Enhanced Attention Network. Built on a ResNet34 backbone, the architecture bifurcates into two specialized streams to process information at different scales:
1. Local CNN Path: Multi-scale Feature Fusion Module (MFFM)
The MFFM is designed to extract fine-grained lesion features. By utilizing pyramid pooling and matrix multiplication, it fuses features from various network depths. This allows the model to integrate high-resolution spatial data from shallow layers with the semantic complexity of deeper layers, significantly improving the detection of subtle, small-scale lesions.
2. Global CNN Path: Dual Attention Module (DAM)
The DAM focuses on broader pathological patterns and the overall distribution of disease. It employs two types of attention:
- Channel Attention: Identifies which feature maps carry the most diagnostic weight.
- Spatial Attention: Directs the network's focus to the most relevant regions of the image, effectively filtering out peripheral noise and artifacts.
The outputs of these two paths are fused before final classification, ensuring the model's decision is informed by both minute details and global context.
Experimental Results and Performance
The model was evaluated using a private dataset from Ningbo Eye Hospital (1,565 images) and the public IDRiD dataset for external validation.
- Classification Accuracy: The dual-path model achieved an overall accuracy of 92.33%, outperforming standard architectures like VGG16 and ResNet50.
- Retinal Detachment (RD) Detection: Notably, the system achieved 100% accuracy in identifying retinal detachment, a condition characterized by large-scale structural changes that the global path is particularly adept at recognizing.
- Generalization: On the IDRiD challenge dataset, the model ranked 2nd overall, demonstrating robust performance even when faced with different image sources and lower input resolutions (512x512).
Conclusion
The integration of local and global processing paths proves essential for the automated analysis of wide-field medical imagery. By effectively managing scale variance and filtering out peripheral noise, the Dual-Path and Multi-Scale Enhanced Attention Network offers a significant step forward in computer-aided diagnosis for ophthalmology. Future work will focus on incorporating transfer learning from standard fundus datasets and addressing geometric distortions inherent in 200° imaging to further refine clinical utility.
Summary updated