Liver Cancer Detection from Multiclass MRI Images using Spatial-Spectral Mamba-Convolution Fusion and Dual-Domain Self-Supervised Graph Attention Networks

Abstract

Liver cancer is one of the major causes of cancer death in the world and requires accurate and automated diagnostic systems based on MRI to intervene at the early clinical stage. Presently, the existing classification schemes of liver cancer are limited by image noise, insufficient discriminative feature learning, lack of contextual representation and lack of optimization capability, which encourages the development of a deep learning-based robust framework for accurate liver cancer detection. In this work, we propose a novel approach called Liver Cancer Detection from Multiclass MRI Images Using Dual-Domain Deep Modular Self-
Supervised Psychologist Skip-Attention Graph Convolution Networks (3DM²S-PSAGCN) for liver cancer detection. First, the Liver Cancer Multiclass MRI Dataset with 5 classes (Angiosarcoma, Cholangiocarcinoma, Healthy liver, Hemangioma,
Hepatocellular Carcinoma) is gathered and resized to 224 × 224. Then, image enhancement is implemented by applying Non-Local Means Denoising and Sparse Dictionary Learning based CNN (NLMD-SDLCNN), which is used to remove noise while maintaining anatomical structures. Then, powerful spatial and spectral representations are captured by the spatial-spectral mamba-convolution fusion network (SSMCFN). Finally, multiclass classification is performed by combining Dual-Domain Self-Supervised Deep Learning with Graph Convolution Network (2D²SDL-GCN) and Deep Modular Skip-Attention Networks (DMSAN), and the network parameters are optimized by using the Psychologist Optimization Algorithm (POA) to achieve better convergence and generalization. Experimental results show the superior performance of the proposed method by achieving an Accuracy of 99.72%, F1-Score of 99.61% and ROC-AUC of 99.88%, which is superior to the existing liver cancer classification methods. The results of this research demonstrate that the proposed framework is successfully tackling the challenges of noisy MRI data, limited feature representation and multiclass discrimination, and can be used to build a reliable and intelligent computer-aided decision-support system, assisting in accurate liver cancer diagnosis.