Publication

Abstract : Liver tumors are a significant medical concern worldwide, with diverse manifestations and varying degrees of severity. Timely and accurately classifying these tumors is essential for effective treatment planning and patient care. In this paper, we present a novel Hybrid model called HybridViT, which combines the power of pre-trained models with the expressive capabilities of Vision Transformers (ViT) for the classification of liver tumors into three distinct classes: Hepatocellular Carcinoma (HCC), Focal Nodular Hyperplasia (FNH), and Hemangioma (HEM). To introduce HybridViT, several pre-trained models are explored, including EfficientNet-B0, EfficientNet-B1, EfficientNet-B7, ResNet-101, ViT B16, and DenseNet-121. Each model is fine-tuned using the liver tumor dataset to evaluate its classification performance. Among the pre-trained models, EfficientNet-B0+ViT achieves a superior test accuracy of 93.71%. This impressive accuracy improvement is attributed to the synergy between pre-trained EfficientNet-B0 features and the ViTs transformer-based approach. The adoption of HybridViT not only advances liver tumor classification but also signifies the broader potential of Hybrid models in medical imaging tasks, presenting new prospects for enhanced diagnosis and treatment of liver tumors.