DefinePK

Abstract

Accurate segmentation and classification of liver tumors are critical for effective clinical diagnosis and treatment planning. While Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) are commonly used modalities for liver imaging, each offers complementary anatomical and functional information. This study presents an attention-based convolutional neural network (CNN) framework for fusing CT and MRI modalities to enhance liver tumor segmentation and classification. The proposed architecture employs dual-branch CNN encoders to extract modality-specific features, which are fused using spatial and channel attention mechanisms for joint representation learning. A U-Net-inspired decoder reconstructs tumor masks for segmentation, while a fully connected classifier predicts tumor type (benign or malignant).
A synthetic multimodal dataset was generated to simulate real-world CT and MRI feature distributions, incorporating segmentation quality (Dice scores) and class labels. The model achieved Dice scores in the range of 0.75–0.92, indicating strong tumor boundary delineation. For classification, the model obtained a macro-averaged F1-score of 0.47 and an AUC of 0.64, demonstrating the potential of attention-guided fusion even under simulated conditions. Attention heatmaps further validated the model’s spatial focus on tumor-relevant regions. These results suggest that multimodal attention-based fusion significantly improves the diagnostic capabilities of CNNs in liver cancer imaging tasks, with promising implications for future clinical deployment.