Search in the Abstract Database

Abstracts Search 2019

P285 Deep learning for automated detection of mucosal inflammation by capsule endoscopy in Crohn’s disease

E. Klang1, Y. Barash1, R. Margalit2, S. Ben Horin2, M. Amitai1, R. Eliakim2, U. Kopylov*2

1Sheba Medical Center, Diagnostic imaging, Ramat Gan, Israel, 2Sheba Medical Center, Gastroenterology, Ramat Gan, Israel

Background

Capsule endoscopy (CE) is a prime modality for diagnosis and monitoring of Crohn’s disease. However, lack of standardisation and prolonged reading time are among the limitations of CE. Recent advancements in artificial intelligence deep learning algorithms present opportunity for utilising this technology in different medical tasks. Utilisation of deep learning techniques may allow for standardised and automated processing of capsule images.

The aim of our study was to evaluate the utility of a deep learning module for detection of small bowel ulcers on CE images.

Methods

We retrospectively collected capsule endoscopy images produced by SB III capsule (Medtronic). Each image was labelled by an expert gastroenterologists either as normal mucosa or containing mucosal ulcers. A state-of-the-art Xception Convolutional Neural Network classified images into either image of normal mucosa or images with mucosal ulcers. The network’s weights were pre-trained on ImageNet data and training was limited to the top fully connected layers. Each capsule image was resized into a 299 × 299 matrix. A fivefold cross-validation, with an 80/20 training/testing split for each fold, was used to evaluate the mean area under the curve (AUC) and accuracy and Youden's index was used to find the models' best sensitivity and specificity for detecting images with mucosal ulcers.

Results

Overall, our dataset included 1363 capsule endoscopy images; 861 normal mucosa images and 502 mucosal ulcers images. Assessment of network training was conducted using plotting of loss and accuracy for training and testing data. The curves of the testing dataset closely follow the curves of testing datasets, which indicates a low degree of overfitting. The mean AUC, accuracy, sensitivity and specificity of the fivefold cross-validation tests for detection of small bowel ulcers were 0.992 ± 0.005, 0.959 ± 0.017, 0.969 ± 0.017, and 0.966 ± 0.023, respectively

Conclusion

Deep learning technology provides highly accurate automated detection of mucosal ulcers on capsule endoscopy CE images. This technology may allow for standardised and automated diagnosis and follow-up of Crohn’s disease by CE in the near future.