Topo-CNN: Retinal Image Analysis with Topological Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Early detection of retinal diseases like diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD) is crucial for preventing vision loss.
• Traditional diagnostic methods for these conditions are often resource-intensive and time-consuming.
• Automated analysis of fundus images offers a promising alternative for efficient and accurate disease screening.

Purpose of the Study:

• To develop an automated and interpretable diagnostic framework for retinal diseases using fundus image analysis.
• To introduce a novel topological feature extraction technique for enhanced performance in disease detection.
• To integrate topological features with deep learning models for a hybrid diagnostic approach.

Main Methods:

• Development of a topological feature extraction method based on Topological Data Analysis (TDA) to capture geometric and structural patterns in fundus images.
• Integration of TDA-derived features with pretrained Convolutional Neural Network (CNN) features (e.g., ResNet-50) into a hybrid deep learning model named Topo-CNN.
• Evaluation of the Topo-CNN model on three public benchmarks: APTOS (DR), ORIGA (Glaucoma), and IChallenge-AMD (AMD).

Main Results:

• Topo-CNN achieved high performance across all evaluated datasets: 98.7% accuracy/98.9 AUC for binary DR, 95.5 AUC for five-class DR, 93.8% accuracy/93.6 AUC for AMD, and 82.3% accuracy/95.8 specificity for glaucoma.
• Ablation studies confirmed the significant contribution of topological features to the model's diagnostic accuracy.
• The proposed Topo-CNN model demonstrated superior performance compared to existing methods on the tested retinal disease datasets.

Conclusions:

• The novel Topo-CNN framework, incorporating topological features, provides an accurate, efficient, and interpretable method for diagnosing major retinal diseases from fundus images.
• Topological Data Analysis offers a valuable approach for extracting meaningful structural information from medical images, enhancing deep learning models.
• This automated diagnostic system has the potential to improve early detection rates and management of sight-threatening retinal conditions.