Real-time image classification for malignant biliary strictures on cholangioscopy images based on deep learning approach

Abstract

It is challenging to determine if the cause of bile duct strictures is benign or malignant. Currently, endoscopists may more precisely inspect the bile duct thanks to computerized single-operator cholangioscopy. As a result, lesions in the bile duct can be seen with the naked eye. However, endoscopists continue to diagnose patients differently. Consequently, a biopsy is typically regarded as the gold standard. The necessity to repeat operations results from a biopsy sample mistake that results in a false-negative cancer diagnosis. In this study, we suggest a convolutional neural network developed particularly for real-time malignant biliary stricture classification. Our approach, which relies purely on an image-level label rather than annotation position, can produce output for both categorization and showing sections of tissue. An augmentation known as "guide-wire augmentation" makes the model focus on tissues rather than equipment, like a guide wire. Our model for still images has been updated to use video inference. All models in our experiment are performed on three patient-based bootstraps. The collection includes 885 images and 104 patient records from King Chulalongkorn Memorial Hospital. The model's sensitivity and F1 performance for still images are 0.8577 and 0.8395, respectively. With a speed of 83 frames per second, the model can be used for real-time inference.