Anomalies Diagnostic in Endoscopic Images Using Deep Learning Ensemble Models

• Pedro da S. Viana UFCA
• Luana B. da Cruz UFCA
• Domingos A. Dias Jr. UFCA
• João Otávio Beira Diniz IFMA

Resumo

Automated anomaly detection in medical images greatly benefits the medical community by assisting in the diagnosis and identification of clinical cases. This work focuses specifically on medical images obtained from endoscopies. We categorized the images into two classes: normal (including normal-cecum, normal-pylorus, normal-z-line) and abnormal (including dyed-lifted-polyps, dyed-resection-margins, esophagitis, polyps, ulcerative colitis). We propose data augmentation to balance the classes and deep learning model techniques for classification. We employed an Ensemble Voting method with deep learning models to enhance the accuracy of the process. Our results are robust, achieving an accuracy of 98.62%, a precision of 96.47%, a sensitivity of 99.64%, and an F1-score of 98.03%. This approach proves to be an effective tool for identifying anomalies in endoscopies. We believe our method can contribute to significant advancements in AI-assisted medical diagnostics.