International Journal of Computer Theory and Engineering

Abstract

Liver cancer ranks sixth among all cancers diagnosed globally. Due to the heterogeneous shape and size of the liver, the manual segmentation of the liver and lesions is a challenging task and time-consuming process. Most of the previous studies in this regard use traditional techniques of image processing to segment the liver and then use handcrafted features to detect lesions and tumors in the liver. The entire process is semi-automatic and results in a loss of information that affects the performance of prediction. Also, deep learning methods employed for liver lesion detection suffer from the misclassification of lesions due to an imbalance of pixel intensities and high processing computational costs. As a result, a new variant U-Net model is designed with a combination of ResNet-18 and ResNet-34 that automatically utilizes 3D contextual information of tumor tissue and detects lesions in the liver. In addition to these, a custom flattened triplet cross entropy function is designed that overcomes the problem of misclassification of lesions due to class imbalance. The novel methodology was evaluated using the benchmark LiTS17 dataset, and the best results were achieved with an accuracy, sensitivity, and specificity of 99.95%, 99.70%, and 99.85%, respectively. We were able to get a considerable reduction in error rate as well as excellent accuracy. The biomedical sector will be transformed as a result of this research.