nnMobileNet: Rethinking CNN for Retinopathy Research.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Convolutional Neural Networks (CNNs) have historically dominated retinal disease (RD) detection.
• Vision Transformers (ViTs) have recently emerged as leading models due to their scalability.
• ViTs' patch-based processing can hinder precise localization of small lesions in RD.

Purpose of the Study:

• To re-evaluate and enhance a CNN architecture (MobileNet) for improved RD diagnostics.
• To investigate if an optimized CNN can rival or surpass ViT performance in RD tasks.
• To provide an accessible, optimized CNN model for retinal disease analysis.

Main Methods:

• Revisiting and modifying the architecture of the MobileNet CNN model.
• Evaluating the optimized MobileNet against ViT-based models on various RD benchmarks.
• Utilizing fundus images for diabetic retinopathy grading, multi-disease detection, and diabetic macular edema classification.

Main Results:

• The optimized MobileNet architecture demonstrated superior performance compared to ViT-based models.
• The CNN model achieved leading results in diabetic retinopathy grading.
• The model also excelled in detecting multiple fundus diseases and classifying diabetic macular edema.

Conclusions:

• Selective modifications to MobileNet can enhance its diagnostic utility for retinal diseases.
• Optimized CNNs, like the revisited MobileNet, can outperform advanced ViT models in specific RD applications.
• The developed model offers a competitive alternative for RD diagnostics, addressing ViTs' limitations in lesion localization.