Many-MobileNet: Multi-Model Augmentation for Robust Retinal Disease Classification

TL;DR

Many-MobileNet introduces a model fusion framework that ensembles multiple lightweight nnMobileNet models with diverse data augmentation strategies to improve retinal disease classification in data-scarce settings. The architecture combines width-scaled CNNs with channel-wise attention, and uses a prediction voting scheme to merge outputs, balancing computational efficiency with robust feature extraction. The authors show robust generalization on ultra-widefield fundus data (UWF4DR) despite limited data, achieving competitive AUROC and AUPRC while highlighting inference-time trade-offs of multi-model fusion. This approach offers a practical path for robust, efficient retinal image analysis in clinical contexts.

Abstract

In this work, we propose Many-MobileNet, an efficient model fusion strategy for retinal disease classification using lightweight CNN architecture. Our method addresses key challenges such as overfitting and limited dataset variability by training multiple models with distinct data augmentation strategies and different model complexities. Through this fusion technique, we achieved robust generalization in data-scarce domains while balancing computational efficiency with feature extraction capabilities.

Figures (4)

• Figure 1: Concept of model fusion during inference.
• Figure 2: Prediction of Many-MobileNet.
• Figure 3: Sample images of UWF4DR dataset, where label 0 represents the ungradable images and label 1 represents gradable images.
• Figure 4: Parameter empirical studies based on UWF4DR dataset.