Learning from small data: Classifying sex from retinal images via deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Deep learning (DL) and convolutional neural networks (CNNs) show promise for automated diagnostics in medical imaging.
• Retinal fundus imaging is suitable for automated analysis, but large datasets are typically required.
• Data privacy and residency restrictions limit the use of massive datasets in clinical settings.

Purpose of the Study:

• To evaluate the performance of DL models on small datasets for classifying patient sex from retinal fundus images.
• To address the challenge of limited data availability in medical AI development.
• To investigate the feasibility of sex classification from fundus images, a previously unquantified trait.

Main Methods:

• Fine-tuning a Resnet-152 model with a modified fully-connected layer for binary classification.
• Conducting experiments on small datasets using private (DOVS) and public (ODIR) data sources.
• Assessing model performance with approximately 2500 fundus images.

Main Results:

• Achieved test AUC scores up to 0.72 (95% CI: [0.67, 0.77]) with small datasets.
• Demonstrated only a 25% performance decrease despite a 1000-fold reduction in dataset size compared to prior work.
• Showcased successful domain adaptation, with models trained on one dataset generalizing to another.
• Highlighted the importance of high-quality images and ensembling for maximizing performance with limited data.

Conclusions:

• Binary classification, including sex categorization from retinal images, is feasible even with very small datasets.
• DL models can generalize well to independent data sources, overcoming distribution shifts.
• Careful data curation and ensembling are crucial for optimizing DL performance in low-data scenarios.
• This study validates the potential of DL for medical imaging applications under strict data constraints.