Auxiliary diagnosis of liver cancer using multimodal CT and MRI with deep learning and radiological feature fusion approach

Liver cancer, particularly Hepatocellular carcinoma (HCC), is a significant cause of cancer-related deaths worldwide due to its asymptomatic nature and challenges in early detection. In clinical practice, radiologists often rely on multiphase Computed Tomography (CT) and multi-sequence Magnetic Resonance Imaging (MRI) to detect and evaluate liver lesions. However, accurately interpreting across these modalities is complex and subject to diagnostic variability, particularly in early-stage cases. Additionally, conventional methods that rely on single-modality CT or MRI often overlook subtle yet critical diagnostic features and fail to integrate cross-modal dependencies effectively. This research presents a Residual Dense-based Multimodal Fusion Network (RD-MFNet) to assist radiologists by providing accurate and consistent auxiliary diagnosis of liver cancer through the combined analysis of CT and MRI scans. The model integrates ResNet3D for modality-specific encoding of CT phases, capturing structural and vascular analysis, and EfficientNet3D with Squeeze-and-Excitation (SE) modules for extracting contrast-enhanced features from image sequences. A residual dense feature fusion using conditional masking handles missing modalities and fuses hierarchical representation across feature phases. The RD-MFNet model enhances liver cancer diagnosis, yielding a 12% increase in Dice Similarity Coefficient (DSC) and a decrease in Hausdorff distance, and achieves a DSC of 92.3% and an area under the curve of 97.5%, demonstrating its reliability in auxiliary diagnosis.