Tu1963 AUTOMATED DIAGNOSIS OF EOSINOPHILIC ESOPHAGITIS FROM LARGE IMAGES OBTAINED BY SPECTRALLY-ENCODED TETHERED CAPSULE REFLECTANCE ENDOMICROSCOPY

Abstract

Eosinophilic esophagitis (EoE) is a chronic inflammatory disease which is caused by food allergies. EoE is manifested by an excessive accumulation of a special type of white blood cells (WBC) called eosinophils within the esophageal wall. The number of eosinophils counted on a histopathologic slide from an esophageal biopsy is the gold standard for EoE diagnosis. Currently the most common threshold for diagnosis is 15 eosinophils/HPF. Since the standard therapy is to remove allergenic foods, patients usually require multiple endoscopies with biopsies to monitor the effect of repetitive dietary controls. This large number of endoscopies places a significant burden on patients and health care system costs. Previously, our group has shown that the eosinophils are highly reflective as seen by bench top reflectance confocal reflectance microscopy (RCM) of fresh human biopsy specimens. This finding motivated our development of a swallowable tethered capsule that conducts high-speed spectrally-encoded reflectance microscopy (SECM) in unsedated patients. The tethered SECM capsule provides cellular-level resolution over the entire esophagus (21 x 200 mm), enabling the scanning of the entire organ for individual eosinophils. One challenge with this technology is the detailed evaluation of the extremely large images (equivalent to 30,000 HPF) to identify regions that contain hypereosinophilia. To mitigate this issue, we trained a convolutional neural network (CNN) with 2,000 labeled images to classify HPF-sized SECM images as either EoE positive or negative. The trained CNN has been validated by predicting 300 images labeled for testing and results demonstrated 85.5% accuracy. These algorithmic and visualization tools have the potential to greatly improve the efficiency of evaluating large-area SECM images of the esophagus.