Deep Learning Approach for Hepatic Lesion Detection

Egypt ranks second worldwide for liver cancer mortality, with 4.57% of global hepatic tumor deaths. Liver disease-related fatalities make up 11.20% of Egypt's total mortality. Additionally, Egypt is among the top 10 countries for obesity, a leading cause of metabolic-associated fatty liver disease. Liver cancer malignancy varies, with the diagnosis and treatment of these differences crucial for patient survival rates and lifespans. Thus, early detection of liver tumors is essential to successful treatment and better patient outcomes. This paper introduces a multiview liver dataset sourced from Ain Shams University Specialized Hospital (ASUSH) in Egypt and annotated by expert radiologists. The dataset includes 280 patients with 14,096 computed tomography (CT) images. This study's ASUSH dataset is groundbreaking, representing Egypt's first multi-class liver cancer dataset with eight distinct classes. Also, furthermore, the paper benchmarks the dataset to multiple deep learning models, including a proposed CNN, as well as fine-tuned pretrained models like VGG16, VGG19, ResNet50, ResNet101, EfficientNetB1, EfficientNetB2, Xception, Vision Transformer (ViT), Inception-V3, and InceptionRestNetV2. Notably, employing an ensemble learning approach significantly enhanced the results for view two. The experimental results on several performance metrics reveal that the models achieved an accuracy of 98% in predicting liver cancer variants.