Use of artificial intelligence to measure colorectal polyp size without a reference object

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Polyp size is critical for determining colonoscopy surveillance intervals.
• Accurate polyp size measurement is essential for effective patient management.
• Current methods may require reference objects, complicating real-time assessment.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) model for measuring colorectal polyp size.
• To assess the AI model's accuracy and clinical utility without a reference object.
• To determine if AI can provide reliable polyp size estimations during colonoscopy.

Main Methods:

• A regression model was developed using SegFormer for segmentation and depth estimation.
• The model was trained on polyp phantoms and fine-tuned with 1,304 real-world colonoscopic images.
• Performance was evaluated on 178 independent images using a snare for ground truth comparison.

Main Results:

• The AI model demonstrated a mean bias of -0.03 mm compared to the snare method.
• Average error rates for polyp size discrepancies were 11.47% across size groups (≤5 mm, 5-10 mm, ≥10 mm).
• High precision (0.870–0.911) and overall recall (0.846) were achieved.

Conclusions:

• AI-based colorectal polyp size measurement is practical and clinically useful.
• The AI model exhibits low error rates and high precision, comparable to traditional methods.
• AI offers a promising, accurate tool for polyp size assessment during colonoscopy without needing a reference object.