Abstract

LIVER tumors pose a substantial global health threat, and in Egypt, they are a significant public health concern. A report in 2020 from the World Health Organization indicated that hepatic tumors contribute to 4.57% ranking Egypt second globally in hepatic tumor mortality and the first in liver disease-related deaths, which constitute 11.20% of total mortality in the nation. It's encouraging to note that within less than a decade, Egypt has drastically reduced its hepatitis C prevalence from 10% to just 0.38%, achieving one of the lowest rates globally by 2023. Recognizing the urgency of accurate diagnoses, the development of computer-assisted imaging techniques using deep learning has become pivotal in recent years. This paper makes noteworthy contributions to the field of artificial intelligence (AI) in the context of medical record digitization. Leveraging AI, especially in medical record systems, has become increasingly essential. The proposed liver dataset, obtained from Ain Shams University Specialized Hospital (ASUSH) in Egypt and annotated by expert radiologists, comprises 280 patients and encompasses approximately 14,096 computed tomography (CT) images sourced from the Picture Archiving and Communication System (PACS). The experiments conducted in two phases are significant. In the initial phase, expert radiologists classify the dataset into binary classification (normal vs. abnormal) cases. Subsequently, in the second phase, the radiologists further categorize abnormal cases into multi-class classifications, these categories include non-tumorous abscesses, malignant metastases, malignant hepatocellular carcinoma treated with radiofrequency ablation, malignant HCC, liver cirrhosis, benign hemangioma, and focal fatty sparing. Notably, this dataset stands out as the first multi-class liver cancer dataset in Egypt, making a novel and impactful contribution to the field. The thesis also benchmarks the dataset using various deep learning models, including VGG16, VGG19, CNN, ResNet50, ResNet101, EfficientNetB1, EfficientNetB2, Xception, Vision Transformer (ViT), Inception-V3, and InceptionRestNetV2. The experimental results are remarkable, with all models achieving an impressive 98% accuracy in predicting liver cancer variants. The evaluation includes several performance metrics, highlighting the robustness and effectiveness of the proposed models in liver cancer diagnosis.

Download

Share

COinS