External Validation of Deep Learning Models for Classifying Etiology of Retinal Hemorrhage Using Diverse Fundus...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal hemorrhage (RH) presents diverse etiologies requiring precise classification for effective patient management.
• Accurate differentiation between traumatic and medical causes of RH is crucial for appropriate clinical intervention.
• Deep learning (DL) models show promise in medical image analysis but require rigorous external validation.

Purpose of the Study:

• To externally validate two deep learning models, FastVit_SA12 and ResNet18, for distinguishing traumatic from medical retinal hemorrhages.
• To assess the performance of these models on diverse, multi-center fundus photography datasets.
• To investigate the interpretability of model predictions using Grad-CAM analysis.

Main Methods:

• Compilation of a comprehensive dataset including private and public fundus image collections (RFMiD, BRSET, DeepEyeNet).
• External validation of FastVit_SA12 and ResNet18 models on a total of 2661 retinal images.
• Performance evaluation using accuracy, precision, and recall metrics, complemented by Grad-CAM for visualization.

Main Results:

• FastVit_SA12 achieved 96.99% accuracy, with high precision (0.9935) and recall (0.9723) for medical RH.
• ResNet18 demonstrated 94.66% accuracy and strong precision (0.9893).
• Grad-CAM analysis revealed distinct activation patterns: ResNet18 focused on global vasculature, FastVit_SA12 on critical regions; medical RH showed localized, traumatic RH diffuse patterns.

Conclusions:

• Both FastVit_SA12 and ResNet18 models exhibit high sensitivity and specificity for classifying retinal hemorrhages.
• External validation enhances the reliability and clinical applicability of AI in ophthalmology for RH diagnosis.
• These validated DL models offer potential for improved patient care and outcomes in managing retinal hemorrhages.