RIA-ViT: A Residual Inception Atrous Vision Transformer for Detection of Gallbladder Cancer from CT Images

Abstract

Gallbladder Cancer (GBC) is a rare but highly aggressive malignancy of the biliary system. In early-stage gallbladder cancer, there are neither overt symptoms nor specific signs. In most patients, it is diagnosed late, resulting in poor outcomes. Even when imaging is available, diagnosing GBC remains difficult because its features overlap with those of non-neoplastic gallbladder pathologies. Furthermore, the overlap between GBC features and those of benign gallbladder conditions may require expert radiological interpretation, which is sometimes unavailable in developing countries. Therefore, the purpose of this study is to design and develop a CT-based framework for automated gallbladder cancer detection that is clinically interpretable, enabling early, consistent, and reliable detection to mitigate diagnostic variability and assist clinicians in making informed decisions. The framework incorporates a hybrid deep learning model, the Residual Inception Atrous Vision Transformer (RIA-ViT), which uses convolutional neural networks to extract local features and vision transformers to contextualize global features. The innovative design incorporates inception residual blocks for multi-scale learning, atrous convolution for enlarging the receptive field, and separable convolution for efficient design. The features are embedded and processed using transformer encoders. The proposed framework was evaluated on a well-curated dataset of CT images of GBC and normal cases, achieving an average classification accuracy of 99.56%. The model accurately separated GBC CT images from normal gallbladder CT images. The proposed framework offers reliable diagnostic outcomes with interpretable, trustworthy visual evidence. The RIA-ViT framework is a major advancement towards the development of automated systems for GBC from CT images.