Multi-Class Liver Cancer Classification using Multiple Deep learning Neural Networks

Abstract

Cancer remains a garnering a prominent position among the deadliest illnesses worldwide. This condition is defined by the uncontrolled proliferation of the body’s cells. Such diseases invade almost all body organs including the liver. Liver cancer can be categorized into primary and secondary types. The primary liver cancer types are cholangiocarcinoma and HCC, both of which are primary liver cancer types. The newest and one of the most prominent types of liver cancer is hepatocellular carcinoma. This cancer type is most associated with the liver and is derived from its various cellular components. Metastasis is a type of secondary liver cancer that occurs when a malignant tumor is formed from any body region and specifically targets the liver. The presence of liver tumors signifies a critical point in cancer evolution; however, the prognosis is much better when liver lesions are detected in the early stages. This study seeks to address the burning issue of early detection and the subsequent treatment of liver cancer. We propose the application of modern deep learning methods to automatically identify lesions deemed cancerous on the liver. This study focuses on the application of Deep Learning, and more specifically, CNNs, and their impact on the accuracy and the quality of diagnosing liver cancer. One of the approaches includes training a CNN to identify primary liver tumors on MRI scans, CTs and Ultrasounds. This study documents the techniques in training the model to detect and contour the liver cancer. This step involves the application of image preprocessing and the training of the neural networks on the extracted features of the liver cancer images... Then, a variety of Deep Learning and other Shallow Learning techniques are used and compared based on system performance to the defined metrics of accuracy, precision, recall and the F-measure. The accuracy of the proposed computer-aided diagnosis system is 95%, the precision is 96%, and the system retains a 95% recall and F1 score. The proposed system outperforms other deep learning models which were ‘fine-tuned’ using ‘transfer learning’.