MRI-Based Liver Cancer Classification with Channel-Spatial Feature Enhancement using ConvNeXt-Tiny with CBAM

Liver cancer continues to be a significant global health issue, with elevated mortality largely attributed to delayed diagnosis and limited treatment options at advanced stages. Accurate classification of liver cancers from Magnetic Resonance Imaging (MRI) plays a crucial role in early detection and clinical decision support. However, conventional imaging assessment and many existing automated approaches often struggle to distinguish subtle variations in liver cancer appearance due to complex tissue characteristics and inter-class similarity. This research presents an attention-enhanced deep learning model based on ConvNeXt-Tiny integrated with a Convolutional Block Attention Module (CBAM) for multiclass liver tumor classification from MRI images. ConvNeXt-Tiny extracts hierarchical spatial features from liver MRI scans, while CBAM refines these features by emphasizing informative channel and spatial responses related to tumor regions. The refined feature representations are subsequently used for classification into five clinically relevant categories: Angiosarcoma, Cholangiocarcinoma, Hemangioma, Hepatocellular Carcinoma and Healthy cases. The model is assessed utilizing a publicly accessible liver MRI dataset, following standard training and testing protocols. Experimental results demonstrate an overall classification accuracy of 98.3%, with consistently high recall, precision and F1-scores across all classes. Receiver Operating Characteristic analysis generated an Area Under the Curve of 0.97, indicating reliable class separation. These findings demonstrate that the developed model supports reliable liver cancer classification from MRI images and offers practical value for imaging-based diagnostic studies.