Predicting the Severity of Liver Cirrhosis with Image Processing Based Machine Learning Models

Abstract

Liver Cirrhosis (LC) is a significant global health concern characterized by a range of liver conditions resulting from chronic alcohol consumption. Over recent decades, LC has contributed to a significant increase in mortality rates worldwide. It states from studies that machine learning and Image processing approaches produce higher accuracy. This study seeks to improve the accuracy of LC detection by developing a robust predictive model. The methodology involves gathering a comprehensive dataset from Kaggle, which includes medical imaging data such as ultrasound, CT scans and MRI scans of both LC-positive and LC-negative patients. Normalization of image dimensions and intensity values ensures dataset consistency followed by liver region segmentation and extraction of texture features like gray level co occurrence matrices (GLCM) to detect abnormalities. Data preparation includes partitioning the dataset into training and testing sets. Utilizing K-nearest neighbors (KNN) imputer for missing value handling, outlier detection methods and addressing class imbalance with Synthetic Minority Oversampling Technique (SMOTE). Feature selection has enhanced through Principal Component Analysis (PCA) and ensemble techniques. The proposed model is evaluated using two machine-learning classifiers: Artificial Neural Network Classifier (ANN) and the Support Vector Machine Classifier (SVM). In the context of LC Prediction, the SVM model outperformed the other model. The ANN model achieved a training score of 0.9738 and a testing score of 0.9478 while the SVM model yielded a training score of 0.9664 and a testing score of 0.9627.The SVM model’s superior testing performance reflects its more reliable real world application. This marks a significant advancement over existing models. In comparison with state of the art techniques, this model's ability to integrate diverse imaging modalities to addressing common challenges like class imbalance and feature selection provides a robust and effective tool for LC prediction. The ability to predict LC severity with high accuracy can greatly enhance early diagnosis. The integration of this model into clinical settings could provide healthcare professionals with a non-invasive diagnostic tool. Future research may focus on incorporating ultrasonography (USS) images into the current model framework to refine diagnostic accuracy. The application will enhance with a more intuitive, user friendly interface and enabling more precise disease diagnosis.