Assessment of Machine Learning Detection of Environmental Enteropathy and Celiac Disease in Children

Key Points Question Can deep learning image analysis distinguish pathological vs healthy features in duodenal tissue? Findings In this diagnostic study, a deep learning convolutional neural network was trained on 3118 images from duodenal biopsies of patients with environmental enteropathy, celiac disease, and no disease. The convolutional neural network achieved 93.4% case-detection accuracy, with a false-negative rate of 2.4%, and automatically learned microlevel features in duodenal tissue, such as alterations in secretory cell populations. Meaning A machine learning–based feature detection platform distinguished pathological from healthy tissue in gastrointestinal biopsy images.


Patient selection
Celiac Disease (CD) and histologically normal control populations were selected from participants who had undergone esophagogastroduodenoscopy (EGD) and biopsy at UVa in the past 20 years. Our criteria for participant sample selection included: 1) age < 6 years for histologically normal controls, and age < 18 years for CD diagnoses, 2) the histopathology report confirmed a diagnosis of CD or reported normal small intestinal biopsy findings, and 3) the corresponding H&E stained slide from the procedure was available at the University of Virginia Biorepository and Tissue Research Facility (BTRF). 76 participants (CD = 34, histologically normal = 42) met our selection criteria and each slide represented all the duodenal samples collected during one EGD.
Participants with EE from Pakistan were part of a larger study that investigated novel biomarkers of EE 1 . This was a prospective birth cohort (n = 380 children) in the rural district of Matiari, Pakistan, where children were followed monthly for growth faltering 1 . Those with a clinical concern for malabsorptive organic pathology (with symptoms such as persistent diarrhea, bloating, etc.) were evaluated by a clinical pediatric gastroenterologist, and diagnostic endoscopies were performed on 11 children. The study case definition of EE included any child with wasting (weight-for-age Z score ≤ -2) found to have weight and height refractory to nutritional intervention 1 . 10 children were found to have EE 1,2 , with 2-3 duodenal biopsies per child for a total of 29 digitized duodenal biopsies available from this subset.
Participants with EE from Zambia were part of another study conducted at the University Teaching Hospital, Lusaka that selected children with severe acute malnutrition 3 . The parent study was designed to investigate the morphological changes and barrier function of the small intestine in children with malnutrition 3 . 34 children with severe acute malnutrition (defined as low weight for height (wasting), i.e. weight-for-height Z score < -2) with persistent diarrhea in whom no cause was identified were hospitalized for endoscopic evaluation 3,4 , any children who were considered too ill for the procedure or had significant congenital abnormalities were excluded 3,4 . Of the children undergoing endoscopy, 16 children were found to have abnormal duodenal histology and were diagnosed with EE. We had access to a single digitized slide per child with EE with each slide having a single duodenal biopsy piece (except for 3/16 cases where there were 2 biopsy pieces) per glass slide.

Background clinical characteristics of patient populations
The characteristics of the EE patients have already been published 2,3 , to summarize the EE patients from Pakistan had a median (Q1, Q3) age of 22 (20, 23) months, 50% were female, 50% (n = 5) had low birth weight (<2500 g), and 60% (n =6) were born premature (< 37 weeks of gestation) 2 . The EE patients from Zambia had a median (Q1, Q3) age of 16.5 (10.5, 21) months, 37.5% were female, and 31.25% were HIV positive. Further details about the characteristics of the patients from Zambia are described by Amadi et al 3 .
Of the 34 CD patients and 42 histologically normal controls, 65% and 38% were female, respectively. The ages of the subsets varied, where the median (Q1, Q3) age for the CD group was 129 (72.5, 180.75) months, and that for the histologically normal control group was 31.5 (22, 49.75) months. LAZ scores could not be calculated for 3 of the CD patients due to missing data, and WAZ scores could only be calculated on 12 of the CD patients as the rest were over the age of 10 years. Of the 31 CD patients, 3 were found to have LAZ < -2. Z scores could be calculated for all of the 42 histologically normal control patients, 5 were found to have LAZ < -2, and 5 were found to have WAZ < -2. Overall, the majority of our patient population from these two groups had appropriate weight and height according to their age.

Deconvolution based categories
Paneth cells play an integral role in our innate immunity, and have been shown to decrease in number in complicated cases of celiac disease 5 . While we were not able to quantify the increase or decrease in Paneth cells per disease category, the deconvolutions predict that they were one of the features used for classification. The deconvolutions also showed detached epithelium and luminal mucin, both of which could be linked to the loss of the epithelial glycocalyx. The glycocalyx is a meshwork formed by glycolipids and/or glycoproteins and protects the intestine by forming a barrier against invading bacteria at the epithelial cell surface 6 . The loss of glycocalyx has previously been observed to be found in the small intestinal biopsies from children with chronic diarrhea and an intolerance to dietary proteins, and it was hypothesized that this may be an immunologic process and not a direct host-protein interaction 7 . EE has commonly been associated with altered barrier function 1,8,9 and we hypothesize that further studying the loss of the glycocalyx could aid in the understanding of the pathophysiology of EE. Further, the detached enterocytes could also serve as a measure of the degree of tissue edema or inflammation. Another interesting group within our deconvolutions was that of "apposed epithelium or enterocytes" -on discussion, we thought this was representative of normal villus architecture, where immediately-adjacent villi would feature apposed epithelial surfaces, while such a feature would be lost in villus atrophy. Both celiac disease 10,11 and EE 2,8,9,12 are known to be associated with villus atrophy, and these convolutions likely indicate that villus structure, specifically villus length and the degree of atrophy, was an important factor for classification.

eFigure 1. Methods of Artificial Data Augmentation
We augmented our data by segmentation, horizontal and vertical reflection of randomly selected patches, and gamma correction. A) Biopsy segmentation; the top image shows automatic segmentation (red boxes), and the bottom image shows manual segmentation; B) Gamma correction; as the biopsy slides were from different centers around the world the H&E stain varied slightly -gamma correction was done to ensure that all the images fed to the deep learning net would not be differentiated on the basis of color but truly on the basis of features present; C) Random selection of patches (blue boxes) and their horizontal and vertical reflections; D) Translation and rotation invariant features; even in different orientations the activations (yellow boxes) corresponded to the same features on each image.

eFigure 2. Feature Predictions t Tests
The activation T-tests serve as a proxy for testing the prevalence of various tissue patterns in different biopsies. In the shown images we are looking at Feature Map 29 which indicates a high activation feature in normal versus celiac and EE (Environmental Enteropathy). This could be a higher presence of goblet cells in normal tissue when compared to celiac and EE or epithelial cell configuration relative to inflammatory cells

eFigure 3. An Illustration of Biopsy Features Correlation Framework Using Data From Pakistan
Our correlation frame works consisted of CNN activations and lasso regression models. Each biopsy was divided, either manually or automatically, into multiple training images which were fed into our CNN model which contained 32 different feature mapping at the fourth layer with each feature map searching for a specific pixel pattern (i.e. mucosal morphological feature) in the input images. Each image, when applied to the CNN framework, produced 25 x 25 convoluted values for each of the 32 feature maps. We then performed a maximum operator on activations across all images per each biopsy. The maximum value of the final 625 convoluted values corresponded to a 142 x 142 segment which maximally activated the feature map. Therefore, each biopsy was mapped onto 32 values that corresponded to 32 segments. We then extracted thirty two 142 x 142 image segments from each biopsy using lasso regression models and correlated their convoluted values at layer 4 with biomarkers. Abbreviations: CNN=Convolutional Neural Network; Max= maximum value.