Performance of a convolutional neural network for automatic detection of blood and hematic residues in small bowel lumen

Capsule endoscopy is most often performed in the setting of obscure gastrointestinal bleeding. Our group has published a Letter to the Editor of Digestive and Liver Disease describing different stages of development of an artificial intelligence algorithm for the automatic detection of blood and hematic residues in capsule endoscopy images.
At its latest stage, this deep learning system demonstrated high sensitivity and specificity (99%), as well as accuracy (98%). Moreover, the convolutional neural network achieved a high image processing performance (186 frames/second). These encouraging results strengthen the belief that the development of accurate AI tools may allow to significantly increase the diagnostic yield of capsule endoscopy and overcome some of its main drawbacks, particularly reading time. You can find the complete article here.

In Press // February 2021