Advanced Residual Transformer (ART) method for identification of bile duct dilatation on CT data

• Published in: Expert systems with applications 2025-03, Vol.266, p.125927, Article 125927
• Main Authors: Güneş, Ali, Kılıçarslan, Serhat, Dönmez, Emrah
• Format: Article
• Language: English
• Subjects: Bile duct dilation; CNN; StyleGAN3; Transfer learning; Transformers
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.125927

Bile duct strictures refer to sections where the intrahepatic or extrahepatic biliary ductal system narrows. This narrowing obstructs the regular forward flow of bile, leading to proximal dilatation and significant morbidity due to the clinical and pathological consequences of biliary obstruction. In this work, we propose a deep learning-based approach to address the detection of bile duct narrowing. Our approach involves using original images as well as generated synthetic data by using StyleGAN3 to train a classification detection network. Through rigorous experiments, we prove that our proposed augmentation strategy not only improves the accuracy of the classification detection model, but also contributes to the stabilization of the training process while significantly reducing the training time. From the experimental results, it is observed that both datasets exhibit high accuracy, with models trained on hybrid dataset achieving exceptional results, surpassing 98 % in most metrics. The accuracy is above 80 % in BBD dataset and above 97 % in hybrid dataset. The study’s main contribution is to compare the performance of Convolutional Neural Network (CNN), Transformer, and GAN techniques, as well as to investigate the effect of detecting biliary dilatation images. The study is the first to compare and successfully apply transformer approaches and GAN techniques to real-world biliary dilatation disease data.