Detection of Barrett's Esophagus in Endoscopic Images Using Machine Learning Algorithms

BE is a precursor to esophageal adenocarcinoma in which the patient is at a higher risk compared to those who the disease is not diagnosed. As for cancer, its early detection and proper handling should be of a significant concern to avoid chances of becoming CANCER. This paper aims at investigating the application and development of machine learning particularly Convolutional Neural Network (CNN) for automated identification of BE when seen in Endoscopy. Utilizing a mixture of open-source databases and specific clinical materials, a complex and well-established CNN model was created and calibrated to diagnose BE. Its efficacy was further supported by high performance rates such as the accuracy of 91% and sensitivity and specificity of 89% and 92% respectively along with AUC ROC value of 0. 94. To solve for the overfitting problem, data augmentation and transfer learning strategies were applied to improve the generalization of the model. Also, to increase the explainability of the deep-learning-based prediction model, certain explainable AI techniques were adopted. The model has been compared with the traditional ML model as well as the basic CNN, and the improved model seems to perform better. About the early detection of BE, this study reveals that the application of the ML-based tools shall provide new ideas and methods to help clinicians to detect BE at its early stage so as to improve the quality of patients’ lives and the survival rate of death from esophageal adenocarcinoma. The clinical photographs generated by this software still require further fine-tuning and refinement to alleviate issues that hinder its implementation in clinical use and to integrate this software seamlessly in clinical settings.