Severity Analysis of Liver Lesions using Enhanced SqueezeNet-Based Feature Extraction with Unsupervised Learning Model

Abstract

Liver cancer remains one of the most lethal malignancies worldwide, necessitating accurate and efficient diagnostic methods for timely intervention. Contrast-enhanced computed tomography (CT) imaging is pivotal in liver cancer screening, with manual segmentation of liver lesions being labour-intensive and prone to variability. In this research, we propose a comprehensive methodology for liver lesion segmentation and severity analysis. Primarily, input images undergo pre-processing steps, including Contrast Limited Adaptive Histogram Equalization (CLAHE) enhancement, to improve image quality. Subsequently, liver segmentation is performed to accurately delineate liver regions, employing the Multi-scale Attention FocusNet (MA-FNet) model specifically designed for liver lesion segmentation. The Unsupervised Learning phase identifies patterns and groupings within the extracted features, which informs the feature extraction followed by severity analysis. A novel Enhanced SequeeNet based Feature Extraction techniques are then utilized to capture relevant information from liver lesions, facilitating unsupervised learning. Our methodology employs a two-step approach for feature extraction and hyperparameter optimization to ensure accurate analysis. Finally, severity analysis provides valuable insights into the extent and severity of liver lesions, aiding clinicians in diagnosis and treatment decision-making. Overall, this research presents a robust framework for liver lesion severity analysis, contributing to improved diagnostic accuracy and patient care in liver cancer management.