Development of an image reading support system for small bowel capsule endoscopy using deep learning

Aim: The aim of this study was to develop a system to support the reading of small bowel capsule endoscopy (SBCE) using deep learning.
Method: In this study, we developed the detection function of abnormal findings by deep learning the image data of 30 cases with abnormal findings. In order to handle a wide variety of abnormal findings, the training data was balanced to include all the major findings identified in SBCE (bleeding, angiodysplasia, ulceration, and neoplastic lesions). In addition, to reduce the false-positive rate, "findings that may be responsible for hemorrhage" and "findings that may require therapeutic intervention" were extracted from the images of abnormal findings and trained. For the performance evaluation, the sensitivity and the number of detections per case were calculated using 271 detectable findings in 35 cases. In addition, the false-positive rate was calculated using 68494 images of non-abnormal findings.
Results: The sensitivity and false-positive rates were 98.4% and 2.8%, respectively. The average number of detection images was 7514.
Conclusion: We developed an image reading support system for small bowel capsule endoscopy using deep learning, and obtained a good detection performance.