Dura-Morphovitnet: Advanced Hybrid Deep Learning for Liver Tumor Segmentation and Classification

In present scenario, liver tumor segmentation and classification from CT scans is critical in analyzing medical images, primarily due to the complexity of liver anatomy and variations in tumor appearance. Traditional approaches struggle with accurate classification of benign and malignant tumors due to limitations in capturing intricate boundary details, spatial consistency differentiation, and lack of sophisticated post-section refinement techniques in modeling both local and global features. To address these issues, this research introduces a novel DURA-MorphoViTNet hybrid deep learning framework, designed for precise liver tumor classification and segmentation. Initially, the liver tumor images are pre-processed with techniques such as rotation, scaling, and elastic deformation applied for enhancing the model’s ability to generalize across variations. The framework uses ResDenseNetCore-MultiVista for feature extraction, combining DenseNet121 and ResNet50 to capture multi-scale features and complex relationships within the data. Then for segmentation, UAttnLabMorphoNet integrates U-Net, DeepLabV3, and scaled dot-product attention to improve boundary refinement and spatial consistency. Further, the post-segmentation refinement is achieved through MorphoCRFNet, which uses Conditional Random Fields (CRFs) and morphological operations to enhance tumor boundary detection, closing small gaps and smoothing irregular edges. Finally, ViT-SpeciCNN combines Vision Transformers and CNNs for multi-class tumor classification, which effectively captures both local details and global dependencies. As a result, the proposed model achieves a noticeable accuracy of 99.92%, precision of 99.76%, recall value of 99.06% and F1-score of 99.02%. This work paves the way for future improvements in liver tumor classification and potential applications in other complex medical imaging tasks.