Toward automated small bowel capsule endoscopy reporting using a summarizing machine learning algorithm: The SUM UP study.

BACKGROUND AND OBJECTIVES

Deep learning (DL) algorithms demonstrate excellent diagnostic performance for the detection of vascular lesions via small bowel (SB) capsule endoscopy (CE), including vascular abnormalities with high (P2), intermediate (P1) or low (P0) bleeding potential, while dramatically decreasing the reading time. We aimed to improve the performance of a DL algorithm by characterizing vascular abnormalities using a machine learning (ML) classifier, and selecting the most relevant images for insertion into reports.

MATERIALS AND METHODS

A training dataset of 75 SB CE videos was created, containing 401 sequences of interest that encompassed 1,525 images of various vascular lesions. Several image classification algorithms were tested, to discriminate "typical angiodysplasia" (P2/P1) and "other vascular lesion" (P0) and to select the most relevant image within sequences with repetitive images. The performances of the best-fitting algorithms were subsequently assessed on an independent test dataset of 73 full-length SB CE video recordings.

RESULTS

Following DL detection, a random forest (RF) method demonstrated a specificity of 91.1 %, an area under the receiving operating characteristic curve of 0.873, and an accuracy of 84.2 % for discriminating P2/P1 from P0 lesions while allowing an 83.2 % reduction in the number of reported images. In the independent testing database, after RF was applied, the output number decreased by 91.6 %, from 216 (IQR 108-432) to 12 (IQR 5-33). The RF algorithm achieved 98.0 % agreement with initial, conventional (human) reporting. Following DL detection, the RF method allowed better characterization and accurate selection of images of relevant (P2/P1) SB vascular abnormalities for CE reporting without impairing diagnostic accuracy. These findings pave the way for automated SB CE reporting.