Diagnósticos de Anomalias Gastrointestinais em Imagens Endoscópicas usando Ensemble Stacking de CNNs e Vision Transformers

Abstract

Automated detection in medical images offers significant benefits, especially in the early identification of clinical cases. This study focuses on images obtained from digestive endoscopes examinations, classifying them into normal categories (normal-cecum, normal-pylorus and normal-z-line) and abnormal categories, which correspond to pathological alterations of the gastrointestinal tract (dyed-lifted-polyps, dyed-resection-margins, esophagitis, polyps and ulcerative colitis). The proposed method encompasses Region of Interest extraction, Specular Highlight, Data Augmentation, and Ensemble Stacking, integrating Convolutional Neural Networks and Vision Transformers. The final model achieved 98.12% accuracy, 98.15% precision, 98.12% sensitivity, 98.23% specificity, and a 98.13% F1-score. These results indicate that the proposed approach has great potential to become an effective tool for identifying anomalies in endoscopic examinations, significantly contributing to AI-assisted medical diagnoses. In addition to increasing the reliability of diagnoses, adopting this system can optimize healthcare professionals’ time, allowing them to focus on more complex and critical cases.