key: cord-0265937-15g9o87o
authors: Berman, A. G.; Fitzgerald, R. C.; Markowetz, F.
title: A composite biomarker for esophageal cancer risk from automated analysis of a non-endoscopic device
date: 2021-08-25
journal: nan
DOI: 10.1101/2021.08.20.21262366
sha: c6585b52e33e07ad6c5f0c46b499755c25424513
doc_id: 265937
cord_uid: 15g9o87o

Barrett's esophagus containing intestinal metaplasia predisposes to cancer, yet the majority of cases are undiagnosed. The length of a Barrett's segment is a key indicator of cancer risk, but measuring it has so far relied on endoscopy, which is expensive and invasive. Cytosponge-TFF3 is a minimally-invasive test that identifies intestinal metaplasia for endoscopic confirmation. We report a machine learning technique to quantify the extent of intestinal metaplasia and predict Barrett's segment length from whole-slide image tile counts automatically generated from Cytosponge-TFF3 histology slides. Utilizing data from 529 patients, our segment length prediction model achieves an average validation fold accuracy of 0.84. Applying this algorithm to an independent test set of 162 patients from a screening trial shows a precision of 0.90 for identifying short-segment disease. This advance will enable higher-risk patients to be prioritized for endoscopy while saving more than half of Cytosponge-TFF3-positive patients from endoscopy in the screening setting.

Of all patients with a TFF3-positive Cytosponge test, 59% had a diagnosis of BE containing To address this challenge, we asked whether it would be possible to use machine learning on 55 Cytosponge-TFF3 to quantify the extent of IM, and correlate that quantity with Prague length 56 measurements ( fig. 1c) . This would provide a composite biomarker of both extent of IM and 57 a segment length estimate from the Cytosponge-TFF3 test alone. This biomarker could have a 58 substantial impact on patient management: patients at higher cancer risk could be prioritized 59 for endoscopy, and those patients likely to have focal cardia IM or a very short segment could 60 be discharged to be followed up with further Cytosponge tests without requiring an endoscopy 61 confirmation.

To develop a composite biomarker, we took the raw output of the deep learning-based After the patient swallows the capsule attached to a thread, the capsule dissolves in the stomach, where the sponge spherical expands. Pulling the sponge by the thread collects cells from the upper stomach and esophagus. The cells are washed off the sponge, processed into a pellet, embedded in paraffin, cut, and stained with TFF3, and put on a slide. Slides are then scanned into whole-slide images (WSIs), which can then be broken up into thousands of smaller images ("tiles"). These tiles are labeled by a pathologist for the presence of goblet cells indicating intestinal metaplasia and a convolutional neural network is trained to perform this classification task from these labels. (b) A diagram showing how the C and M lengths of the Prague classification criteria for segment length measurement are found. C (circumferential) denotes distance from the proximal margin of the gastric cardial folds to the proximal margin of the circumferential BE segment, and M (maximal) describes the distance to the most proximal extent of the segment. GOJ = gastroesophageal junction. 

. CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted August 25, 2021. ; https://doi.org/10.1101/2021.08.20.21262366 doi: medRxiv preprint assay as assessed by deep learning showed a highly significant correlation with segment length. 

As a control, we took the raw output of the Cytosponge-H&E model from (7) for identifying 73 tiles containing gastric tissue but no TFF3-positive tiles. As the presence of gastric tissue in a 74 Cytosponge sample merely indicates that the device traversed through the lower esophageal 75 sphincter into the stomach, it is expected that there should be little to no correlation between 76 these gastric tile count and the Prague C and M lengths. Indeed, for the same 529 BEST2 77 patients, Spearman's rank correlation coefficients between these tile counts and the C length 78 was −0.024, and for M length was −0.046, confirming the null hypothesis for the control 79 setting in comparison to IM tile count. 80 We next asked if the significant correlation between automatically identified IM quantity 81 and the Barrett's segment length could be leveraged in an optimized prediction model. We In the regression-based approach, we found that a zero-inflated Poisson (ZIP) regression 90 model (17) yielded the best fit for the data (see Methods). We performed 5-fold cross validation 91 on our 529 BEST2 patient cohort, all of whom underwent a Cytosponge followed by an oral 92 5 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted August 25, 2021. ; https://doi.org/10.1101/2021.08.20.21262366 doi: medRxiv preprint were not promising for clinical application.

In the classification-based approach, which takes advantage of established clinical thresh-97 olds, we trained a logistic regression model to classify whether or not a patient's Barrett's seg-98 ment had a C length at or above 1 cm or an M length at or above 3 cm (C≥1 or M≥3 vs. C<1 Cytosponge after an interval (e.g., three years) would seem a good alternative to endoscopy.

Our results show that not only is there a correlation between intestinal metaplasia quantifica-128 tion from the Cytosponge-TFF3 test and Barrett's segment length, but that a clinically relevant 129 segment length cutoff is predictable with high accuracy from this quantification. This estab- 

In standard pathology reporting from glass slides it has not been feasible to quantify the 134 degree of IM previously. However, since our predictions are gleaned from an entirely automated 135 machine learning pipeline, this suggests that the incorporation of these new biomarkers into 136 existing Barrett's diagnosis and monitoring pathways should be readily achievable.

There is a big push towards more systematic screening for Barrett's esophagus to improve 138 outcomes from esophageal cancer (8, 12-14) . In doing so, every effort must be made to max-139 imize the benefits and reduce the harms of screening. Hence, clinical investigations that ensue 140 from a positive screening result should spare patients at very low risk for cancer and be af-141 fordable for the health care system. If Cytosponge was adopted for screening on a population 142 7 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted August 25, 2021. ; class-balanced accuracy, precision, recall, and F1 results of 5-fold cross-validation on 529 BEST2 patients applied to a logistic regression model for predicting whether or not the patient's Barrett's segment had C length greater than or equal to 1 cm or M length greater than or equal to 3 cm. L ("long segment") stands for the C≥1 or M≥3 class; S ("short segment") stands for the C<1 and M<3 class. (h) Results of training a model like that of (g), except training on all 529 BEST2 patients without cross-validation and then inferring on a 162-patient test set of BEST3 patients. BEST3 results shown. 8 . CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. 

. CC-BY-NC 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted August 25, 2021. ; https://doi.org/10.1101/2021.08.20.21262366 doi: medRxiv preprint

Global cancer statistics 2018: Globocan estimates of incidence and mortality world-187 wide for 36 cancers in 185 countries

Barrett oesophagus

Evaluation of a minimally invasive cell sampling device coupled with 194 assessment of trefoil factor 3 expression for diagnosing barrett's esophagus: a multi-center case-195 control study

Role of tff3 as an adjunct in 197 the diagnosis of barrett's esophagus using a minimally invasive esophageal sampling device-the 198 cytosponge

Triage-driven diagnosis of barrett's esophagus for early detection of esophageal 202 adenocarcinoma using deep learning

Dysplasia and cancer in a large multicenter cohort of patients with barrett's esoph-204 agus

Risk of malignant progression in barrett's esophagus patients: Results from a large 206 population-based study

The development and validation of an endoscopic grading system for barrett's 208 esophagus: the prague c & m criteria

Acg clinical guideline: Diagnosis and 210 management of barrett's esophagus

Development and validation of a model to determine risk of progression of barrett's 220 esophagus to neoplasia

Asge guideline on screening and surveillance of barrett's esophagus. Gastroin-222 testinal Endoscopy