Identification of eye diseases with small fundus image datasets using a new unsupervised deep learning approach

This is the authors' abstract. We don't add key points for this paper.

Ocular diseases are a severe global problem, particularly in countries that lack the technology or economy to treat them. Automated analysis of retinal fundus images could enable early screening, but supervised deep learning methods require large, annotated datasets that are expensive and time-consuming to obtain. This study introduces an unsupervised deep learning framework for learning and clustering of retinal image representation that achieves strong performance on small-scale fundus image datasets using only image-level disease labels to guide contrastive pair construction and classification loss, without requiring pixel-level annotations for clustering. The proposed method, NF-HAE-UDL (normalizing flow-guided hybrid autoencoder-based unsupervised deep learning), combines a hybrid autoencoder architecture (ResNet-50 and Swin Transformer) with normalizing flow and contrastive learning to learn discriminative latent representations. The autoencoder generates meaningful feature representations, upon which the normalizing flow maps the encoded features to a more expressive and structured latent space. Meanwhile, contrastive learning encourages the separation of patterns targeted to specific diseases through data augmentation. Evaluation on the STARE dataset (81 images, multi-label) and the IDRiD subset (211 images, binary) demonstrates superior unsupervised clustering performance compared to traditional methods, achieving ARI scores of 0.58 ± 0.10 on STARE and 0.58 ± 0.06 on IDRiD. Auxiliary classification validation on frozen encoder representations confirms that learned embeddings preserve disease discriminative information (88% in-sample accuracy on STARE, 98% on IDRiD), though these results represent quality assessment rather than diagnostic performance. Furthermore, Grad-CAM visualizations validated against expert annotations (IoU = 0.52 ± 0.11 ) highlight clinically meaningful attention to pathological features. The framework demonstrates a promising direction for effective, interpretable, unsupervised retinal image analysis with limited data.