The best reference standard for size measurement after polypectomy so far is caliper-based size measurement. However, this is time-consuming and may be inaccurate because of polyp fragmentation, distortion, or other deformations due to piecemeal resection, retrieval-related damage, inappropriate compression, or folding. This underscores the need for novel measurement tools.
MAGENTIQ-COLO (MAGENTIQ EYE, Haifa, Israel) is an AI-based system that offers real-time polyp size assessment during colonoscopy or retrospective video evaluation. A video-based validation study was performed to assess the accuracy of this AI system compared to caliper-based size measurement of colorectal polyps.
A total of 77 polyps was measured in 55 patients (male sex, n (%): 29 (52.7%), age median, interquartile range (IQR): 67 (13)). Colonoscopy indications were surveillance in 38 (69.1%), screening in 7 (12.7%), and diagnostic in 10 patients (18.2%). Of the 77 polyps, the majority of polyps were adenomas (n=52; 67.5%). Polyps were equally distributed through the colon. The AI-based system correctly identified size in 37/46 polyps (80.4%) as ≤5 mm and 5/6 polyps (83.3%) as ≥10 mm (Table 1). The system had a mean accuracy of 0.80 for ≤5 mm polyps (n=46), 0.88 for 6-9 mm polyps (n=25), and 0.83 for ≥10 mm polyps (n=6), with a total mean accuracy of 0.84. Accounting for the skewed number of polyps per size category, the weighted accuracy was 0.83, with a weighted sensitivity and specificity of 0.83 and 0.92, respectively. The system reliably estimated polyp size, with a slightly higher accuracy for >5 mm polyps compared to ≤5 mm polyps. From initial polyp detection to showing size, median latency was 5 frames (±0.17 seconds).
Confusion matrix of ground truth polyp size by caliper and predicted polyp size by the AI-based system
|
|
AI system (≤5 mm) |
AI system (6-9 mm) |
AI system (≥10 mm) |
|
Caliper (≤5 mm) |
37 |
9 |
0 |
|
Caliper (6-9 mm) |
1 |
22 |
2 |
|
Caliper (≥10 mm) |
0 |
1 |
5 |
The novel AI-based system demonstrated a high accuracy compared to the reference standard of calipers, signifying that the AI system correctly identified most polyps within their true size categories with a low risk of size misclassification. Furthermore, colonoscopy time only slightly increased, allowing timely size assessment. This system could well replace the caliper as the reference standard for in-vivo size measurement of polyps during colonoscopy, but further prospective studies are needed.
Acknowledgement: We express our gratitude to the MAGTENIQ-COLO study group for their contribution to drafting this abstract.