Abstract
BACKGROUND: Liver cancer, including hepatocellular carcinoma (HCC), is a leading cause of cancer-related deaths globally, emphasizing the need for accurate and early detection methods. OBJECTIVE: This study introduces LiverCompactNet, an advanced deep learning framework for the early detection and classification of liver cancer. METHODS: LiverCompactNet classifies liver images into three categories: benign, malignant, and normal. The dataset comprised 5,000 liver images (1,500 benign, 1,500 malignant, and 2,000 normal), divided into training (3,500), validation (750), and test (750) subsets. Data preprocessing involved normalization using MinMaxScaler, class balancing. Additionally, exploratory Principal Component Analysis (PCA) was performed only on derived tabular features (e.g., intensity histograms, categorical encodings) to visualize variance structure, but PCA was not directly applied to raw imaging data or CNN training inputs. RESULTS: LiverCompactNet demonstrated outstanding performance with an overall accuracy of 99.1%, malignant detection sensitivity of 98.3%, specificity of 99.4%, precision of 97.6%, and an AUC-ROC score of 0.995. Training performance steadily improved, with accuracy rising from 90% in epoch 1 to 99% by epoch 20, and validation accuracy increasing from 88% to 98.5%. Loss analysis revealed effective learning, with training loss approaching zero and validation loss remaining marginally higher. Final evaluations confirmed near-perfect classification metrics: precision at 97.6%, sensitivity at 96.8%, specificity at 98.9%, and an AUC-ROC score of 0.993. CONCLUSION: LiverCompactNet offers highly accurate, reliable, and early detection capabilities for liver cancer, paving the way for improved medical image analysis and clinical decision-making.