Automated Multi-class Liver Cancer Classification from MRI Using a Cross-gated NASNet–vision Transformer Architecture

Liver cancer is a predominant cause of cancer-related death globally. Timely and accurate diagnosis is
required for improving survival and treatment results. Magnetic resonance imaging (MRI) is pivotal in the
identification and clinical evaluation of liver cancer. Reliable liver cancer classification is challenging due to
heterogeneous tumor morphology, subtle inter-class variations among different cancer subtypes, and low contrast
between malignant, benign, and healthy liver tissues. Conventional diagnostic approaches rely heavily on handcrafted radiomic features and radiologist-driven interpretation, which are limited by inter-observer variability, imaging complexity, and scalability in large-scale screening scenarios. This study presents a hybrid Neural Architecture Search
Network (NASNet) –Vision Transformer (ViT) model with Cross-Gating Fusion for automated multi-class liver cancer
classification from MRI images. The NASNet component extracts discriminative local spatial and textural features,
while the vision transformer modules capture long-range global contextual and inter-channel dependencies. The crossgating fusion mechanism adaptively integrates complementary representations, enhancing feature discrimination and
classification robustness. The proposed model was evaluated on a liver MRI dataset comprising 11,002 images across
five classes. Experimental results demonstrate excellent performance, achieving an overall classification accuracy of
98.54%, along with good precision, F1-score and recall values across all liver cancer categories. The model exhibits
stable convergence, robust generalization capability, and minimal misclassification, outperforming recent CNN-based
and radiomics-driven approaches. These findings underscore the efficacy of integrating convolutional and transformerbased representations via cross-gated fusion, providing a robust and scalable solution for clinical decision support
systems and MRI-based liver cancer diagnosis.