Deep learning-based detection of eosinophilic esophagitis

Abstract

Background: For eosinophilic esophagitis (EoE), a substantial diagnostic delay is still a clinically relevant phenomenon. Deep learning-based algorithms have demonstrated potential in medical image analysis. Here we establish a convolutional neuronal network (CNN)-based approach that can distinguish the appearance of EoE from normal findings and candida esophagitis.

Methods: We trained and tested a CNN using 484 real-world endoscopic images from 134 subjects consisting of three classes (normal, EoE, and candidiasis). Images were split into two completely independent datasets. The proposed approach was evaluated against three trainee endoscopists using the test set. Model-explainability was enhanced by deep Taylor decomposition.

Results: Global accuracy (0.915 [95 % confidence interval (CI) 0.880-0.940]), sensitivity (0.871 [95 %CI 0.819-0.910]), and specificity (0.936 [95 %CI 0.910-0.955]) were significantly higher than for the endoscopists on the test set. Global area under the receiver operating characteristic curve was 0.966 [95 %CI 0.954-0.975]. Results were highly reproducible. Explainability analysis found that the algorithm identified the characteristic signs also used by endoscopists.

Conclusions: Complex endoscopic classification tasks including more than two classes can be solved by CNN-based algorithms. Therefore, our algorithm may assist clinicians in making the diagnosis of EoE.