A Comparative Analysis of Erythematous Mucosa Identification in the Small Bowel Using Machine Learning Methods

Identification of gastrointestinal (GI) tract abnormalities, such as erythematous mucosa, not only indicates the presence of underlying inflammation but also aids in the decision-making process of treatment. Wireless Capsule Endoscopy (WCE), a noninvasive GI diagnostics modality, is utilized for abnormalities identification before possible intervention procedures. Due to its focus-intense and time-consuming process, manual examination of WCE recordings impacts gastroenterologists' day-to-day duties and leads to potential oversight of G I abnormalities. Machine Learning (ML) methods have the potential to play a pivotal role in automatically detecting and identifying G I abnormalities with their advanced capabilities. This work aims to assess the performance of base and ensemble ML methods to detect and identify erythematous mucosa in WCE recordings. Two experiments were conducted on a local WCE dataset, with 166 labeled frames (erythematous mucosa) and 2185 frames (normal mucosa). Several evaluation measures were calculated, including, accuracy, precision, recall, F1-score, misclassification rate, Matthews Correlation Coefficient, and Cohen's Kappa. As a result, ensemble methods demonstrate superior performance over base methods, achieving a perfect accuracy rate of 100 %. This research supports the clinical deployment of ML-driven frameworks for the detection of abnormal GI tissues, opening the door for more accurate diagnostic instruments in clinical settings.