Deep learning architectures for abnormality detection in endoscopy videos

Abstract

Endoscopy is a widely employed technique for the diagnosis and treatment of various internal organs in the human body, including the gastrointestinal tract, lungs, bones, and abdominal region. During the procedure, an illuminated optical device records video data, which assists physicians during real-Time analysis and post-procedure evaluations. Identifying areas of interest within the vast amount of recorded video data is critical for optimizing physicians' focus and time. A key task in this process involves classifying endoscopic frames as normal or abnormal. Current solutions for endoscopic frame classification either rely solely on handcrafted features or neural network features and lack efficient pre-processing techniques to eliminate irrelevant frame portions or enhance relevant region features. This study presents an innovative architecture pipeline for the efficient and robust detection of abnormal frames in endoscopic videos, combining effective pre-processing techniques with deep neural networks. A novel and customized pre-processing method has been integrated into three custom-Tailored deep architectural pipelines, which are based on sequential convolutional networks, InceptionResNet, and EfficientNet. Models generated using these pipelines were trained and tested on customcurated data from publicly available repositories. Among the three pipelines, the architecture based on EfficientNet outperformed current state-of-The-Art approaches, achieving a sensitivity, specificity, and accuracy of 0.94, 0.91, and 0.93, respectively, for the classification of abnormal frames. This novel approach demonstrates the potential of leveraging advanced deep learning architectures to enhance abnormality detection in endoscopic videos.