Vision-language large learning model, GPT4V, accurately classifies the Boston Bowel Preparation Scale score

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Gastroenterology

Background:

• Large learning models (LLMs) are advanced AI, initially for natural language processing, now adapted for multi-modal tasks.
• Boston Bowel Preparation Scale (BBPS) scoring is a clinically relevant task for assessing colon cleansing.
• Traditional AI requires extensive data; this study explores LLMs' potential with limited examples.

Purpose of the Study:

• To evaluate the efficacy of a vision-language LLM for automated Boston Bowel Preparation Scale (BBPS) grading.
• To determine if LLMs can achieve high accuracy in BBPS classification with fewer training examples compared to traditional AI methods.

Main Methods:

• Utilized GPT4V, a vision-language LLM from OpenAI, via API.
• Employed a standardized prompt with contextual references for BBPS grading.
• Tested performance on the HyperKvasir dataset for automated BBPS classification.

Main Results:

• GPT4V achieved 98% valid results on 1794 images.
• Accuracy was 0.84 for two-class (BBPS 0-1 vs 2-3) and 0.74 for four-class (BBPS 0, 1, 2, 3) classification.
• Macro-averaged F1 scores were 0.81 and 0.63, respectively, comparing favorably to traditional methods.

Conclusions:

• Vision-language LLMs can accurately perform BBPS classification without large training datasets.
• This demonstrates a paradigm shift for AI in medicine, particularly for conditions with limited data.
• LLMs offer a promising approach for AI classification in data-scarce medical scenarios.