Cholangiocarcinoma, a form of liver bile duct cancer, is challenging to detect due to its critically low 5-year survival rate. Conventional imaging modalities, such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), are widely used, but recent advancements in Hyperspectral Imaging (HSI) offer a promising, noninvasive alternative for cancer diagnosis. However, supervised learning methods often require large annotated datasets that can be difficult to obtain. To alleviate this limitation, we propose an unsupervised learning strategy using Generative Adversarial Networks (GANs) for cholangiocarcinoma detection. This approach, named Unsupervised Spectral and Spatial Attention-based GAN (USSGAN), employs an unsupervised Spectral-Spatial attention-based GAN to classify and segment cancerous regions without relying on labeled training data. The integration of an adaptive step size into Tasmanian Devil Optimization (TDO) enhances the convergence speed and effectively captures diverse cancerous features. Enhanced Tasmanian Devil Optimization (ETDO) further improves segmentation performance, making the framework robust and computationally efficient. The proposed method was tested on a publicly available multidimensional choledochal cholangiocarcinoma dataset, achieving superior performance compared with existing techniques in the literature. USSGAN demonstrated high accuracy across key metrics such as overall accuracy (OA), average accuracy (AA), and Cohen’s Kappa. Ablation studies confirmed the critical contributions of the proposed enhancements to the overall performance. With an overall accuracy of 98.03%, the USSGAN closely aligns with the assessments of experienced pathologists while maintaining minimal computational requirements. Its lightweight nature ensures real-time deployment, providing results within a minute, making it a practical and effective solution for clinical applications.