Detection of Liver Cancer in CT Images using Machine Learning Techniques

Abstract

newlineLiver cancer is one of leading cause of death all over the world. Medical image analysis is an important part in the process of diagnosis oftheliver cancer. Clinical research have shown that computed tomography (CT) scans play amajor role in diagnosing liver cancer at early stages. It helps the oncologists in the process of diagnosis process and treatment planning. Manual detection process regularly done to detect liver cancer by medical experts, which is time consuming and expensive. Hence development of automated methods using computer-aided diagnosis (CAD) is necessary for accurate detection of liver cancer. newlineIn this thesis, five different algorithms have been proposed for automatic detection and classification liver cancer using CT scan images. The frame work for detection process consists of four steps: segmentationof liver, extraction of cancerous lesion, feature extraction and classification. Separation of liver from other organs is done using adaptive thresholding with watershed transform. The cancerous part in the liver is segmented using different methods such as Kernelized and spatial fuzzy C-means, level set, optimized techniques and Gaussian mixture model. These techniques are validated on a large datasets consists of 225 real-time CT scans taken from different subjects affected by liver cancer at IMS and SUM Hospital, India. The informative features are extracted using Gray-Level Co-Occurrence Matrix (GLCM), local binary pattern (LBP) descriptor and wavelet transform methods to form a large dataset. The feature set is then classified into different types of cancer i.e. hemangioma (HEM), hepatocellular carcinoma (HCC) and metastatic carcinoma (MET) using neural network (NN), support vector machine (SVM), random forest (RF) and deep neural network (DNN) classifiers. The performance of the proposed techniques are evaluated in terms of sensitivity, specificity, accuracy, and Jaccard index. It has been observed that watershed Gaussian based deep learning (WGDL) techniques have effectively detected the liver cancer and DNN classifiers achieved best classification result having an accuracy of 0.99 with Jaccard index of 0.98 at 200 epochs with a negligible validation loss of 0.062. Hence the proposed methodology of segmentation and use of machine learning techniques has potential for the use in diagnosis of liver cancer in clinical practices.