Vision transformer-based feature selection for liver tumor MRI analysis

Liver tumors, classified as benign or malignant, pose significant diagnostic challenges. In infancy, benign liver tumors may progress to malignancy, making early and accurate classification crucial. Traditional manual classification methods are inefficient, time-consuming, and prone to errors, necessitating advanced automated techniques. This study introduces a novel Vision Transformer with Learned Invariant Feature Transform-based statistical features (ViT+LIFT based Stat features) approach for liver tumor classification. Magnetic Resonance Imaging (MRI) liver tumor images from the ATLAS dataset serve as input. The preprocessing stage employs an Adaptive Wiener Filter (AWF) to enhance image quality. A Dynamic Context Encoder Network (DCE-Net) is then utilized to segment the liver and lesions. Feature extraction incorporates Shape Index Histogram (SIH), shape features, ResNet features, and LIFT with statistical features. Finally, the Vision Transformer (ViT) classifies liver tumors based on these extracted features. The proposed ViT+LIFT based Stat features model achieved superior classification performance, with an accuracy of 91.732%, sensitivity of 90.118%, and specificity of 90.710%. These results demonstrate the effectiveness of the proposed method in improving liver tumor classification accuracy, reducing diagnostic delays, and minimizing the need for invasive biopsies.