S1311 CNN Analysis of Video Capsule Endoscopy Frames Reduces Miss Rate of Relevant Lesions

Abstract

INTRODUCTION: Human reads of small bowel video capsule endoscopy (VCE) have a surprisingly high miss-rate for known lesions.1 We previously reported development of a convolutional neural networks (CNN) with very high sensitivity and specificity for detection of lesions on VCE images.2 We aimed to determine if this CNN can detect relevant lesions missed by human readers of VCE. METHODS: A total of 146 VCE studies (not previously exposed to CNN) were evaluated for this retrospective study. Each video was de-identified and analyzed by the CNN to select frames predicted to contain abnormalities. The frames saved by the original reader (‘OR’) were verified to include examples of all relevant lesions mentioned in the de-identified report. Abnormalities identified in ‘OR’ frames and CNN frames were reviewed and classified by 3 experts as clinically relevant, possibly relevant, or innocent. Each abnormality found by ‘OR’ and CNN were aligned by timestamp to determine if found by CNN only, ‘OR’ only, or by both CNN and reader. RESULTS: From these 146 VCEs, there were a total of 542 abnormal findings (17% were relevant, 47% were possibly relevant and 36% were innocent (Table 1). Of the definitely relevant findings, 68% were found by both the ‘OR’ and the CNN, 1% was found by only the ‘OR’ (missed by CNN), and 31% were found by only the CNN (missed by ‘OR’). Of the possibly relevant findings 36% were found by both, 5% were found by only the ‘OR’, and 36% were found by only the CNN. Of the innocent findings 55% were found by both, 19% were found by only the ‘OR’, and 58% were found by only the CNN. Broken down by lesion type, miss rates of definitely relevant lesions by ‘OR’ vs CNN are shown in Table 2. Of the 94,000 mean number of frames analyzed per video, the CNN generated a mean of 200 frames with abnormalities, of which, a mean of 31 were confirmed to have abnormalities (Table 3). CONCLUSION: This retrospective comparison of lesion detection by original reader vs CNN in 146 VCE studies suggests that use of CNN may significantly reduce miss rate of clinically relevant lesions, especially AVMs, bulges, and polypoid lesions.Table 1.: Relevance of abnormalities found reader vs CNNTable 2.: Miss rates among "Definitely Relevant" lesionsTable 3.: Total frames generated by VCE and CNN compared to frames with abnormalities