DEEP LEARNING BASED AUTOMATIC POLYP SIZE ESTIMATION IN REAL-TIME DURING COLONOSCOPY

Aims Accurate polyp size estimation (PSE) during colonoscopy is essential to determine appropriate resection methods and surveillance intervals. Automatic PSE (APSE) may help to reduce PSE variability through standardization. The MAGENTIQ-COLO is an AI system that includes real-time polyp detection and real-time APSE. The purpose of this study was to assess APSE performance.

Fig. 1

Methods APSE categorizes polyps between three groups,<5mm (62 polyps in the dataset), 5-10mm (20 polyps), and>10mm (9 polyps). It is implemented by a regression Convolutional Neural Network (CNN) with a backbone of ResNet101 CNN, and was trained on 672,533 frames from 795 polyps. The results are displayed in real time for detected polyps. ([Figure 1]) A total of 58,771 frames of 91 polyps taken from 79 colonoscopy videos were used. We measured the accuracy, precision, and sensitivity of each size group and overall. These measurements were done frame-wise, comparing the APSE against the endoscopists’ annotations in procedure reports. These annotations were based on estimates the endoscopists made with forceps or snares during polypectomy.

Results Overall APSE accuracy was 90.0%, overall APSE sensitivity was 81.9%, and overall APSE precision was 86.1%. The accuracy, sensitivity, and precision of the APSE calculated for each polyp size group is displayed in [Table 1].

Table 1

Metric/Group	<5mm	5-10mm	>10mm	Overall (average)
Accuracy	88.91%	85.15%	95.89%	90.0%
Sensitivity	87.23%	85.78%	72.66%	81.9%
Precision	90.44%	77.38%	90.49%	86.1%

Conclusions The newly introduced APSE allows for accurate prediction of colon polyp size in real-time. APSE allows for a more accurate and standardized diagnosis with less variability compared to current PSE techniques, and in this way assists in correctly determining colonoscopy surveillance intervals.

Publication History

Article published online:
14 April 2022