Quantitative measurements of crypts from advanced endoscopy imaging using deep learning-based segmentation

The quantitative measurement of intestinal crypts morphology and their architecture can provide valuable insights on the normal intestinal functioning and on subtle changes linked to disease onset and progression. To ensure accurate measurement, the identification and accurate delineation of the crypts boundaries is paramount for the extraction of morphometrics features that then need to be linked to clinical assessment. Endocytoscopy enables the virtual histology and real time assessment of mucosa with its high magnification capability to visualize the crypts and other microstructure. We used a Mask Region-based Convolutional Neural Network (Mask R-CNN), to localize and segment crypts, to extract the quantitative morphometric index which may potentially be useful to characterize patients with healthy and inflamed mucosa in Ulcerative Colitis (UC). A total of 65 endocytoscopy videos from 47 patients are processed for segmentation followed by quantitative measurements of the crypts. For the Mask R-CNN segmentation process, on test data of four patients, no false positive is detected, and sensitivity observed is 94% with overall accuracy of 96%.
Then, from all frames with crypts segmentation, we extract parameters such as crypt density, area, eccentricity, diameter, average distance of the crypt from the neighboring crypts. We observed that the automatic measurements have a 95% correlation with those extracted from the manually annotated crypts.