Dense Inception V5 Convolution Neural Network for Liver Tumor Classification into multi abnormal instances and staging of the Disease using LiTS-CT images

Abstract

Liver Cancer is one among the deadliest form of the cancer disease which owing to abnormal development of cells in the liver and its surrounding tissue. Hence identification and classification of the liver lesion or malignant growth through manual observation is highly challenging due to complex boundaries and features with high degree of intraclass variation and low degree of interclass variations. Machine learning based unsupervised algorithm has been employed to automatically classify the diseases on basis of the lesion appearance and its characteristics but those models consume more processing time and will lead to reduced scalability and reliability. In order to tackle those limitations, deep learning architecture has been exploited as it is more advantageous in discriminating the liver lesions features efficiently and accurately in order to prevent cancer cells from multiplying and spreading. In this paper, a novel dense Inception V4 Convolution Neural Network for liver cancer classification and staging of the disease on processing of the CT images has been proposed. Initially selective median filter and contrast limited adaptive histogram equalization on CT images employed as preprocessing technique to improve the results of the image segmentation through noise removal, contrast enhancement and normalization of the images. Next, region growing segmentation has been employed to pre-processed images to segment the region of the interest and lesion boundaries effectively. Those segmented image has been employed to principle component analysis which acts feature extraction technique to extract the normalized multiple lesion feature of the liver cancer region. Extracted feature has been employed to the model of the generate the learning model on the employing Dense Inception V4 Convolution Neural Network Classifier for label smoothing on disease classification with 7*7 convolutions on optimizing the  hyperparameter for filter vector outputs. Further proposed model minimize the complexity of the network and enhances the computing efficiency. Experimental results of the proposed model have been evaluated in the MATLAB software on using LiTs dataset. Performance analysis of the proposed model 3 classes of the disease as basal hepatocellular carcinoma, hemangioma and liver metastasis with 98.75% accuracy, 98.46 specificity and 99% sensitivity respectively on comparing against conventional classifiers