Decoding the Best AI Models for Retinal Screening

Deep learning is carving a significant niche in the area of automated retinal screenings. Despite the progress, the challenge remains: how do various visual model families stack up in multi-disease scenarios and domain shifts?

The Contenders

In a comprehensive benchmark using the Retinal Fundus Multi-disease Image Dataset (RFMiD), twelve architectures across four model families were scrutinized. The candidates included convolutional neural networks (CNNs), vision transformers, hybrid CNN-transformer backbones, and vision-language models.

The focus was on two tasks: binary screening for any retinal disease and a more nuanced multi-label classification spread across 28 distinct disease classes. Standardized training, calibration, and evaluation protocols ensured a level playing field, with metrics such as AUC, F1, precision, recall, and sensitivity providing critical insights.

Performance Matters

On the RFMiD, all architectures delivered commendable results in binary screening, boasting an AUC above 84%. However, attention-based models took the lead. The SwinTiny and hybrid CoAtNet0 and MaxViTTiny models not only excelled in binary screening but also improved macro and micro F1 scores in the multi-label setting. The question is, are these models setting a new standard for clinical deployment?

Meanwhile, vision-language models like CLIP ViT-B/16 and SigLIP-Base384 proved competitive against CNN baselines, yet they didn't quite surpass the top-performing transformers and hybrids. It's a reminder that the container doesn't care about your consensus mechanism. performance is what counts.

The Bigger Picture

In an external validation on Messidor-2 for referable diabetic retinopathy, AUC scores ranged from 66.8% to 84.7%. Here, too, hybrid and transformer models showcased strong performance. It's clear that these results provide a reproducible reference for model selection in multi-disease retinal screening and pave the way for future automated screening tools in clinical settings.

The takeaway? While all modern deep learning models bring something to the table, attention-based models are carving out a distinct advantage. Enterprise AI is boring. That's why it works. The ROI isn't in the model. It's in the 40% reduction in document processing time. As the medical field continues to embrace AI, the choice of model can have significant real-world implications for patient care and operational efficiency.