Grounded report generation for enhancing ophthalmic ultrasound interpretation using Vision-Language Segmentation...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Ophthalmic ultrasound interpretation is vital but requires expertise and time.
• Increasing ultrasound data necessitates efficient AI solutions for analysis and reporting.
• Existing AI models lack lesion identification and interpretability.

Purpose of the Study:

• To introduce the Vision-Language Segmentation (VLS) model for ophthalmic ultrasound interpretation.
• To enhance diagnostic accuracy and interpretability in AI-driven ultrasound analysis.
• To evaluate the VLS model's performance and cost-effectiveness.

Main Methods:

• Developed the VLS model by integrating a Vision-Language Model (VLM) with the Segment Anything Model (SAM).
• Trained and tested the model on 64,098 images and 21,355 reports from three hospitals.
• Evaluated performance using BLEU4 scores, dice coefficients, specificity, and diagnostic accuracy.

Main Results:

• Achieved high BLEU4 scores (66.37 internal, 85.36/73.77 external) and dice coefficients (59.6% internal, 50.2%/51.5% external).
• Demonstrated strong diagnostic accuracy (90.59% internal, 71.87% external) with high specificity (97.8%/97.7%).
• Showcased a 30-fold cost reduction in report generation, from $39 to $1.3 per report.

Conclusions:

• The VLS model offers an interpretable and accurate solution for ophthalmic ultrasound analysis.
• This AI approach significantly reduces manual effort and accelerates diagnostic workflows.
• VLS presents a promising advancement for efficient and cost-effective ophthalmic diagnostics.