A Machine Learning Framework for Robust Detection Of Healthy And Ulcer Tissue In Endoscopic Images

Abstract

In clinical endoscopy, rapid and reliable tissue classification between healthy and ulcerated regions is crucial for effective diagnosis and treatment planning purposes. This study proposes a novel computational pipeline that integrates deep feature extraction with classical machine learning to discriminate healthy and ulcer tissues from endoscopic images. A pre-trained EfficientNetB0 model was used to extract high-level image features, which were then reduced using supervised Partial Least Squares (PLS) and classified with a Radial Basis Function Support Vector Machine (RBF-SVM). With the incorporation of class imbalance mitigation strategies, such as controlled oversampling and probability reweighting, the framework demonstrates strong performance and interpretability. The effectiveness of the proposed method is demonstrated through an evaluation of a curated endoscopy image dataset, achieving high accuracy and robustness in distinguishing tissue types. The results indicate its potential for clinical decision support in gastroenterology.