An interpretable artificial intelligence system for measuring the size of small polyps (< 10 mm)
Yong Li1, Fujun Li2, Guanghui Lian1
1Gastroenterology, Xiangya Hospital Central South University, China.
Background:
Precise polyp size measurement is vital for treatment and follow-up strategy determination, with small polyps (<10 mm) presenting particular challenges. This study aimed to develop an artificial intelligence system, PolypM, for automated polyp size measurement to enhance clinical decision-making.
Methods:
PolypM, comprising two models, was designed for automatic segmentation of transparent caps and polyps. It was trained on 6486 endoscopic images (Dataset 1), validated on 675 images (Dataset 2), and compared to endoscopists' measurements on 542 images (Dataset 3).
Results:
The PolypM trained on Dataset 1 achieved an intersection over union (IoU) of 0.91 [95% confidence interval (CI): 0.89-0.93] for segmenting transparent caps, an IoU of 0.75 (95% CI: 0.71-0.79) for segmenting polyps, and an intraclass correlation coefficient (ICC) of 0.682 compared to the gold standard in Dataset 2. The PolypM also demonstrated comparable accuracy on Dataset 3. In the multicenter external validation, the PolypM outperformed endoscopists in both average absolute error and average relative error in determining polyp size (P<0.05) regardless of polyp morphology, pathological characteristics, or endoscopists' experience.
Conclusions:
PolypM, an interpretable system for colonic polyp size estimation, was developed to mitigate measurement ambiguity and facilitate decision-making regarding surgical intervention and postoperative surveillance timing.
