ChatOCT: Embedded Clinical Decision Support Systems for Optical Coherence Tomography in Offline and Resource-Limited...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Optical Coherence Tomography (OCT) is vital for diagnosing eye and systemic conditions.
• Accessibility is limited by specialized expertise and high computational needs.
• Existing systems often require internet connectivity, hindering use in resource-limited areas.

Purpose of the Study:

• Introduce ChatOCT, an offline, domain-adaptive Clinical Decision Support System (CDSS) for OCT.
• Enhance accessibility and utility of OCT diagnostics in diverse clinical settings.
• Develop a resource-efficient AI solution for real-time medical image analysis.

Main Methods:

• Utilized LLaMA-2-7B, incorporating domain knowledge from PubMed and OCT News via two-stage training (knowledge injection and Q&A instruction tuning).
• Implemented activation-aware weight quantization to reduce GPU memory usage to ~4.74 GB for offline deployment.
• Developed a novel expert answer generation framework to mitigate hallucinations and ensure response accuracy.

Main Results:

• ChatOCT achieved 10-15 point higher scores in coherence, relevance, and clinical utility compared to baselines.
• Reduced GPU memory requirements by 79% while maintaining real-time responsiveness (~20 ms inference time).
• Expert ophthalmologists validated ChatOCT's outputs as clinically actionable and aligned with diagnostic needs.

Conclusions:

• ChatOCT is an innovative offline CDSS for OCT, deployable on local embedded hardware for real-time analysis.
• The system offers a scalable, generalizable pipeline for resource-efficient clinical AI solutions.
• ChatOCT has the potential to significantly assist frontline healthcare providers, especially in resource-limited settings.