PERFORMANCE OF A CONVOLUTIONAL NEURAL NETWORK FOR THE DETECTION OF BLOOD OR HEMATIC RESIDUES IN ENTEROSCOPY: A PROOF-OF-CONCEPT STUDY

Aims Artificial intelligence algorithms have shown promising results when applied to different endoscopic techniques. The application of Convolutional Neural Networks (CNN) for detection of lesions in double-balloon enteroscopy (DBE) has not been explored. We aimed to develop and test a CNN-based algorithm for automatic detection of blood or hematic residues in DBE exams.

Methods We included a total of 6900 images, 1435 showing blood or hematic residues. The remaining images showed normal mucosa or other findings. A pool of 5520 images (80% of the image dataset) was used for development of the network. Its performance was evaluated using a validation dataset comprised by the remaining 20% of the dataset (n=1380). The output provided by the network was compared to a consensus classification provided by two gastroenterologists with experience in DBE ([Figure 1]). The sensitivity, specificity, accuracy, positive and negative predictive values, and area under the curve (AUC) were calculated.

Results After optimization of the neural network, our model automatically detected blood in the small bowel in enteroscopy images with a sensitivity of 95.8%, a specificity of 97.6%, positive and negative predictive values of 91.4% and 98.9%, respectively. The CNN had an overall accuracy of 97.2%. The AUC was 0.99. The CNN analyzed the validation dataset in 10 seconds, at a rate of approximately 138 frames per second.

Fig. 1

Conclusions We developed a pionner AI algorithm for automatic detection of blood or hematic residues during DBE exams which may enhance the diagnostic yield of deep enteroscopy techniques in patients with bleeding originating from the small bowel.

Publication History

Article published online:
14 April 2022