Article Preview
Top1. Introduction
According to the World Health Organization (WHO), liver cancer is the second most common type of cancer. Several types of cancer can form in the liver, of which hepatocellular carcinoma (HCC) is the primary type, which starts in the hepatocyte. Intrahepatic cholangiocarcinoma and hepatoblastoma are other types of liver cancer; however, they occur less frequently. The classification of liver cancer is based on texture and volume, and this helps in the meticulous planning of liver therapy. Authors in (Cornelis et al., 2017) explains classification techniques is performed based on two types of livers: the first one, hypodense, appears darker than the healthy liver, and other one is brighter.
The appearance is based on lesion type, imaging, and state and can vary among patients. In this regard, the accurate classification of liver tumors is paramount in medical imaging.
Computer-aided diagnosis through machine learning and computer vision is mainly developed for doctors and patients to make timely decisions. However, there is a lack of research on the analysis of different architectures of convolutional neural networks (CNNs) and the deployment of solutions to the cloud for public use.
The research conducted so far has focused on wavelet coefficient statistics (Kumar et al., 2013) and grey level co-occurrence matrix, (Chang et al., 2017; Gunasundari et al., 2018; Sayed et al., 2016), which are used for the extraction of shape and texture features in liver CTs. For tumor classification in the liver, artificial computer-aided diagnosis and support vector machines (SVMs) are used as existing techniques, but the accuracy achieved so far is nominal. To achieve greater accuracy, GoogLeNet architecture is used and a 3D construction of imaging is employed in our existing work. Further, ImageJ/Fiji is used for 3D construction and analysis. Supported by NIH Image for the Macintosh, ImageJ is an open-area Java image processing tool that runs either as a downloadable application or as an online applet on any PC with a Java 1.4 or later virtual machine. It can investigate, process, show, alter, spare, and print 8-bit, 16-bit, and 32-bit pictures. It can also peruse many image formats including TIFF, BMP, DICOM, GIF, JPEG, FITS, and “raw.” It supports “stacks,” a progression of pictures that share a single window.
Additionally, the program has been deployed on a cloud platform. Microsoft Azure is used for cloud deployment, and the software is utilized as a remote desktop application. In many fields, deep learning is a significant research topic, i.e., CNNs are used to address the challenges in computer vision. CNN offers outstanding performance in terms of handwritten character recognition, object recognition, and image classification. Here, supervised learning models for multi-layer networks are introduced by the authors in (LeCun et al., 1998). CNN can retrieve hierarchical features without any manual intervention by constructing high-level features. Several medical applications use CNN for liver tumors (Li et al., 2015) and colonic polyps (Guo et al., 2017). Fig. 1 illustrates the human liver.
In our research, liver tumor segmentation and classification are carried out using CNN by considering training images from the TCIA repository. The training model is created with respect to GoogLeNet, SqueezeNet, and AlexNet and accuracy is checked using all three models. Further 3D models are constructed using the ImageJ Tool, where liver image identification is performed without a loss in its structure. Finally, the solution is deployed in the Microsoft Azure cloud as a remote desktop application, and similar processes can also be deployed in local servers so that applications can be accessed at higher speeds. A training model is generated considering the images chosen manually from the TCIA repository.