A Smart MSqueeze-Link Model for Robust Liver Tumor Classification Using Advanced TransU-Net Segmentation Process and CT Images

Abstract

Due to the liver's intricate anatomy, interpreting Computed Tomography (CT) images to diagnose liver disease can be demanding for experienced radiologists. Since determining tumor features, including their type, size, and level of progression, often proves difficult, interest in Computer-Aided Diagnostic (CAD) methods has surged, as these systems play a significant role in enhancing early detection and treatment, thereby extending patient life expectancy. This study developed a hybrid Modified Squeeze-LinkNet (MSqueeze-Link) strategy for classifying liver tumors in CT imagery, which operates through four stages: preprocessing, segmentation, feature extraction, and classification. In preprocessing, the CT image is refined through a Gaussian filtering technique. This is followed by segmentation, during which an Enhanced Transformer U-Net model (E-TUNet) precisely separates tumor structures from surrounding tissue. Next, feature extraction is performed using Improved Pyramid Histogram of Oriented Gradients (I-PHOG), color-based measures, and deep features from AlexNet, Residual Network (ResNet), and Visual Geometry Group 16 (VGG16). The final classification stage combines the strengths of ModifiedSqueezeNet (MSqueeze) and LinkNet, as each model processes the extracted features independently, and their outputs are given to a two-dimensional Global Average Pooling (2D GAP). The pooled outputs are concatenated and forwarded to a dense layer to produce the final tumor classification result. Experimental results on a public dataset show that the proposed MSqueeze framework surpassed previous methods in segmentation and classification.