Subscribe to RSS
DOI: 10.1055/s-0042-1744609
A COMPUTER-ASSISTED AUTOMATED APPORACH FOR OPTICAL CLASSIFICATION OF COLORECTAL POLYPS INCLUDING SERRATED ADENOMAS – THE CASSANDRA STUDY
Aims Computer-assisted models (CAM) aim to differntiate neoplastic and non-neoplastic polyps based on their optical features. However, differentiation of serrated adenomas (SA) from hyperplastic polyps (HP) and adenomas (AD) is still challenging. We aspired to develop a CAM for automated polyp classification between said polyp classes.
Methods Polyps of 250 patients were resected. Histological diagnoses were used as reference standard. A total of 489 videos of 327 polyps were recorded. Of these, 191 videos were used for CAM development. CAM corresponds to a region proposal deep neural network based on the RetinaNet architecture trained for recognizing the three classes (Figure1). After development, 100 new polyps were presented to CAM in order to test the program. The 100 test polyps were also presented to two experts (E1, E2). We compared CAP-based and human accuracy. Primary endpoint of the study was CAM-based accuracy.
Results
Factor |
Computer |
Expert 1 |
Expert2 |
---|---|---|---|
Sensitivity for SA |
6.7% (1/15) |
60.0% (9/15) |
60.0% (9/15) |
NPV for SA |
85.9% (85/99) |
90.8% (59/65) |
91.9% (68/74) |
Specificity for SA |
100.0% (85/85) |
69.4% (59/85) |
80.0% (68/85) |
PPV for SA |
100.0% (1/1) |
25.7% (9/35) |
34.6% (9/26) |
CAM-based accuracy regarding the prediction of SA was 86.0%. Sensitivity and negative predictive value (NPV) were 6.7% and 85.9%. CAM-based accuracy of SA prediction was higher compared to E1 (86.0 vs. 68.0%, p=0.004). For adenoma prediction CAM-based accuracy, sensitivity and NPV was 57.0%, 96.0% and 81.8%. Experts accuracies for adenoma prediction surpassed CAM-based accuracy (p=0.001 respectively). Inter-rater agreement of optical predictions was good between both experts (72% agreement k=0.58).
Conclusions An automated, computer assisted differentiation of SA from HP or AD is feasible. However, differentiating three different polyp classes seems to pose challenges to the CAM approach. More video data is needed in order to refine the CAM.
Publication History
Article published online:
14 April 2022
© 2022. European Society of Gastrointestinal Endoscopy. All rights reserved.
Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany