Development and External Validation of Novel Deep Learning Platforms for the Diagnosis and Risk Stratification of Dysplasia in Barrett Esophagus Histological Whole-Slide Images
Abstract
This Fiscal Year 2022 Peer Reviewed Cancer Research Program Impact Award application is responsive to the Esophageal Cancer Topic Area and the Prevention and Diagnostics Military Health Focus Area. Esophageal cancer is a lethal cancer with less than 20% survival in Veterans, when diagnosed after the onset of symptoms. Barrett s esophagus is the only known precancerous condition, giving rise to esophageal cancer via the development of dysplasia (precancerous cellular changes seen in biopsy tissue before the onset of cancer). BE is present in 8%-25% of Veterans undergoing endoscopy. Dysplasia can be detected on biopsies taken from the esophageal lining via endoscopy. Hence, periodic endoscopy is recommended to detect dysplasia in BE patients. Endoscopic biopsies are evaluated by pathologists (doctors with specialized training in the assessment of tissue samples). Dysplasia is graded as low grade (LGD) or high grade (HGD) and increases the risk of developing esophageal cancer. Non-surgical endoscopic treatments can treat dysplasia and prevent the development of esophageal cancer in Barrett s patients with dysplasia. Hence, correctly diagnosing precancerous changes (dysplasia) in endoscopic biopsies from patients with Barrett s esophagus is critical, to enable timely treatment to prevent esophageal cancer. Unfortunately, the diagnosis of dysplasia in BE biopsies is challenging, with several subjective criteria used for evaluating dysplasia (not present, low grade, high grade). Consequently, pathologists often do not agree with each other on the dysplasia grade. This variation and inconsistency, leads to either inappropriate treatment or lack of treatment when it is required. This problem is more common when pathologists do not have experience in reading Barrett s pathology slides. This mostly happens in smaller hospitals. A proposed solution is that biopsies in which precancerous change (dysplasia) is diagnosed by less experienced pathologists in smaller hospitals should be re-checked by expert pathologists with more experience (usually in larger university hospitals). Additionally, access to experts is limited, criteria to define an expert, and for grading dysplasia severity are lacking or subjective. To address these important knowledge and access gaps, we have developed an artificial intelligence (AI) tool for the accurate diagnosis of precancerous change in BE slide images. This tool was developed at the Mayo Clinic using images of tissue samples from almost 600 patients and is very accurate, compared to Mayo Clinic expert pathologists. As next steps to further develop this model before widespread use, we are proposing this study to: (1) Test the accuracy of this AI model in slide images of 680 patients who will be taking part in an ongoing national National Institutes of Health-funded trial of BE patients with LGD. The grade of precancerous change (dysplasia) will be confirmed in these patients by a panel of expert pathologists at the Cleveland Clinic. This aim will enable the model to be tested on a large sample of non-Mayo Clinic biopsy images, which were not used to develop it. The model may also be adjusted to make it more accurate using additional slide images from this trial. (2) Test the accuracy of this AI model in slide images of 450 patients from community pathology practices (from all parts of the United States), sent for confirmation of precancerous changes to Mayo Clinic. The results of the model will be compared to the diagnosis made by a panel of four national expert GI pathologists (one from Cleveland Clinic, one from University of Miami, two from Mayo Clinic). This aim is important to enable implementation of this model in community pathology practices. Additionally, the impact of this model on the confidence of pathologists in making a diagnosis of precancerous changes on slides will also be tested. (3) Build and test an AI model to predict future development of esop
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2024
- Source ID
- HT94252311079
Entities
People
- Prasad Iyer
Organizations
- Mayo Clinic
- United States Army