Compact vision language models enable efficient and interpretable optical coherence tomography through layer-specific...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate clinical narratives from optical coherence tomography (OCT) B-scans are crucial for diagnosing retinal diseases.
• Current algorithms struggle to integrate visual features with domain expertise for effective interpretation.

Purpose of the Study:

• To develop an efficient AI model for generating clinical narratives from OCT scans.
• To improve the accuracy of disease classification in retinal imaging.

Main Methods:

• A multimodal dataset of 40,000 OCT B-scans was curated, paired with expert-validated summaries.
• Introduced LO-VLM, a compact vision-language model (VLM) with anatomical guidance for summary generation and classification.
• Benchmarked LO-VLM against RetinaVLM, LLaVA-Med, and a ViT model.

Main Results:

• LO-VLM narratives achieved a mean score of 8.5/10 in blinded specialist evaluations, significantly outperforming RetinaVLM (5.5/10).
• Achieved 80.3% SBERT similarity and 71.5% BERTScore F1, surpassing specialized VLM baselines.
• Reached 96% accuracy in disease classification, outperforming ViT by 13% and medical VLMs by over 62%.

Conclusions:

• LO-VLM offers a paradigm for efficient and interpretable AI models in OCT interpretation.
• The model successfully reconciles computational efficiency with high performance in clinical narrative generation and disease classification.