Tumor attention networks: Better feature selection, better tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Traditional computed tomography (CT) scan analysis for liver tumors has limitations in precision and inter-observer variability.
• Convolutional Neural Networks (CNNs) show promise in medical image segmentation but struggle with imitating human visual attention and feature selection.
• A need exists for more effective and interpretable CNN architectures for liver tumor segmentation.

Purpose of the Study:

• To introduce a novel network architecture, Tumor Attention Networks (TA-Net), for improved liver tumor segmentation.
• To enhance feature selection mechanisms in CNNs by embedding adaptive Tumor Attention layers.
• To validate the effectiveness and interpretability of TA-Net for hepatic disease diagnosis and treatment monitoring.

Main Methods:

• Developed TA-Net, a novel network architecture incorporating Tumor Attention layers with multi-functional modules.
• Employed 3D and 2D perspectives within tumor attention layers to adaptively highlight tumor features and suppress irrelevant ones.
• Conducted extensive experiments, including ablation studies and evaluations on multiple datasets, to assess performance and generalization.

Main Results:

• TA-Net demonstrated superior liver tumor segmentation performance compared to state-of-the-art methods across various metrics.
• The proposed network achieved this improvement with lower computational cost and minimal parameter overhead.
• Visualization results confirmed the effectiveness and interpretability of the tumor attention modules.

Conclusions:

• TA-Net offers a significant advancement in automatic liver tumor segmentation, providing precise tumor volumes and reducing variability.
• The network's adaptive feature mining capabilities enhance diagnostic accuracy and assist in therapeutic decision-making for hepatic diseases.
• TA-Net shows strong generalization capabilities across different medical image datasets and segmentation tasks.