Compact Vision-Language Models Enable Efficient and Interpretable Automated OCT Analysis Through Layer Specific...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Interpreting Optical Coherence Tomography (OCT) B-scans for retinal diseases requires AI that combines visual data with medical knowledge.
• Existing AI models struggle to accurately translate OCT imaging features into clinical narratives.

Purpose of the Study:

• To introduce LO-VLM, a novel, compact vision-language model (VLM) designed for OCT B-scan interpretation.
• To enhance free-form summary generation and multiclass disease classification of retinal conditions.

Main Methods:

• Curated a multimodal dataset of 40,000 OCT B-scans with expert-validated summaries for six conditions.
• Developed LO-VLM, a 247M parameter VLM incorporating anatomical guidance in its encoder and decoder.
• Benchmarked LO-VLM against RetinaVLM, LLaVA-Med, and a ViT model.

Main Results:

• LO-VLM narratives received significantly higher scores (8.5/10) from retina specialists compared to RetinaVLM (5.5/10).
• Achieved superior quantitative metrics: 0.803 SBERT similarity and 0.715 BERTScore F1.
• Reached 96% accuracy in disease classification, outperforming ViT by 13% and medical VLMs by over 62%.

Conclusions:

• LO-VLM offers a new paradigm for efficient and interpretable AI in OCT analysis.
• The model demonstrates superior performance in both generating clinical narratives and classifying retinal diseases.
• LO-VLM reconciles computational efficiency with high accuracy for OCT interpretation.