TA-Net

TA-Net segments liver tumors in computed tomography (CT) scans using Tumor Attention layers to improve tumor feature selection and enable precise tumor volume estimation.

Key Features:

• Tumor Attention layers: Adaptive attention modules highlight tumor-related features and suppress irrelevant information to enhance segmentation.
• 2D and 3D feature mining: The architecture extracts and integrates features from both 2D and 3D perspectives of CT data.
• Multi-functional modules: Built-in modules improve interpretability and effectiveness in liver tumor segmentation tasks.
• Skip connection configurations: Various skip connection designs are explored to optimize information fusion across the network.
• Robust feature selection: Addresses inconsistencies in conventional multi-layer and multi-kernel convolutional neural networks (CNNs) by providing an explicit feature selection mechanism.
• Performance and efficiency: Demonstrates higher liver tumor segmentation accuracy with lower computational cost and minimal parameter overhead compared to state-of-the-art methods.
• Benchmark validation: Experimental evaluation on clinical benchmark data and additional medical image datasets assesses performance and generalization.
• Comparative analyses: Evaluated against general semantic segmentation methods and non-tumor segmentation tasks to characterize versatility.
• Tumor volume estimation and variability reduction: Provides precise tumor volume estimations aimed at reducing inter-observer variability.

Scientific Applications:

• Liver tumor segmentation: Automated delineation of liver tumors in CT scans for quantitative analysis.
• Tumor volume estimation: Quantification of tumor size to support therapeutic decision-making and monitoring of treatment efficacy in hepatic diseases.
• Generalization to medical imaging: Application and validation across multiple medical image datasets to assess cross-dataset robustness.
• Method benchmarking: Comparative benchmarking with semantic segmentation and non-tumor segmentation approaches.

Methodology:

TA-Net employs a novel network architecture with Tumor Attention layers and multi-functional modules, integrates 2D and 3D feature mining, explores various skip connection configurations, and is evaluated via comparative analyses on clinical benchmark data and additional medical image datasets.