Improving Liver Cancer Diagnosis: A Multifaceted Approach to Automated Liver Tumor Identification in Ultrasound Scans

This research paper introduces a comprehensive methodology aimed at enhancing the accuracy of liver cancer tumor identification in ultrasound images through the utilization of a Support Vector Machine (SVM) classifier. The proposed approach encompasses four key steps, commencing with a novel noise-filtering technique that adapts the "peak-and-valley" method through pixel scanning along the Hilbert curve. Subsequent steps involve the implementation of a "Windows adaptive threshold" procedure to further refine noise reduction and optimize Otsu's algorithm for precise segmentation threshold determination. The identification of disconnected objects is facilitated through the "core area" method, supported by a feature knowledge base. The final step involves employing the SVM classifier to categorize liver images into normal and tumor classes. A significant advancement in impulsive-like noise reduction is introduced through the novel nonlinear noniterative multidimensional filter—the peak-and-valley filter. Rooted in order statistics, this filter minimizes reliance on background information and efficiently replaces noisy pixel values using conditional rules. Its multidimensional versions allow for recursive reduction, presenting independent filters applicable in preferred sequences. Comparative analysis with the widely used median filter demonstrates competitive performance, with the peak-and-valley filter proving notably faster, positioning it as a promising replacement in sophisticated filtering methods. The research comprehensively explores diverse liver tumor types, including "Abscess," "Cyst," "Hemangioma," and "Hepatocellular carcinoma," each characterized by unique features crucial for differentiation. Analysis of area, centroid coordinates, boundary box dimensions, orientation, and min/max gray-level values across these tumor types forms the foundational elements for SVM classifier training. Otsu's method seamlessly integrates into the proposed solution for optimal threshold selection, capitalizing on its nonparametric and unsupervised nature, simplicity, and applicability to multi-thresholding problems. The SVM classifier, renowned for robust nonlinear classification, operates independently of feature space dimensionality and excels with limited training samples. The testing phase on representative liver images demonstrates the model's potential efficacy in accurately identifying normal and tumor liver images based on distinctive features. This contributes to the early and precise diagnosis of liver conditions, showcasing the significance of our multifaceted approach. The research establishes a robust and efficient framework for automatic liver cancer tumor identification, amalgamating enhanced image processing techniques and SVM classification to improve diagnostic accuracy in non-invasive liver cancer diagnosis, with an accuracy of 91.8 ± 4.2%. The associated standard deviation of 4.2% suggests a moderate level of variability, indicating dependable and consistent performance across diverse scenarios.