Multi-constraint Deep Inception-V7 DenseNet 169Architecture for Liver Lesion Classification and Tumor Staging

Abstract

Liver Cancer is one of the deadliest cancer types which propagating around the world due to its increased morbidity and mortality. Complications associated with liver cancer is considered as fibrosis, cirrhosis, obesity, smoking, and hepatitis B and C. Liver cancer is high challenging to be diagnosed at an initial stage due to reduced primary symptoms and it initiate in the deep location of the body without leaving any symptoms. Machine learning models has been increasingly implemented for monitoring disease prognosis and for diagnosing initial disease. However, those models are time consuming and error prone due to critical features extraction on the segmented disease regions. Henceforth, new architecture for early identification of liver cancer becomes mandatory to identify the liver deteriorations and manage the disease from further liver deterioration. In order to manage those complications, a new deep learning prototype represented as Multi-constraint Deep Inception V3 DenseNet169  is designed for liver lesion classification and tumor staging. Initially acquired data is preprocessing using wiener filter as it produces the enhanced image quality with high contrast and sharpness for efficient segmentation of the tumor regions.  Next, Yolov7 algorithm is applied to pre-processed images to segment the lesion and non legions region with efficient placement of boundaries effectively. Those segmented image has been applied further to feature extraction technique named as Non-linear discriminant analysis to extract the multiple lesion feature of the liver lesion segments. Extracted feature from lesion segments has been employed to the proposed deep learning model mentioned as multiparameter Inception V7 DenseNet169 Classifier of the produce the liver lesion classification and staging of disease on basis of its feature scores. DenseNet 169 layer architecture is considered as classifier which is composed of  convolution layer for feature mapping of the extracted features, max pooling layer to extract the high level features map and fully connected layer utilizes the softmax function on employing naive bayes classifier to classify the feature map into liver lesion classes such as hepatocellular carcinoma, hemangioma and liver metastasis .Further current architecture is capable to reduce the challenges of the network and increases the computing efficiency using loss function named as cross entropy. Experimental outcomes of the current architecture is accessed using MATLAB software on using LiTs dataset which contains 1500 CT images. Performance analysis of the current architecture produces the liver disease classes such as cirrhosis, fibrosis, basal hepatocellular carcinoma, hemangioma and liver metastasis with 98.65% accuracy, 97.46 specificity and 98.84% sensitivity respectively on compared with state of art deep learning and machine learning classifiers