ALA-Net: Adaptive Lesion-Aware Attention Network for 3D Colorectal Tumor Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of colorectal tumors in 3D MRI is crucial for clinical applications like staging and radiotherapy.
• Existing methods struggle with noise, misclassification, and incomplete segmentation due to insufficient feature aggregation and loss of spatial information.

Purpose of the Study:

• To develop a novel network, the adaptive lesion-aware attention network (ALA-Net), to address limitations in 3D medical image segmentation.
• To improve the accuracy and reliability of colorectal tumor and tissue segmentation on 3D MRI.

Main Methods:

• Proposed ALA-Net utilizes two parallel encoding paths to capture global context and fine spatial details.
• Incorporated a lesion-aware attention module to adaptively capture long-range dependencies and enhance discriminative features.
• Introduced a prediction aggregation module for multiscale feature fusion and precise voxel-wise prediction.

Main Results:

• ALA-Net demonstrated superior performance compared to state-of-the-art methods in 3D medical image segmentation.
• The network achieved improved target completeness and a reduction in false positives.
• Accurate detection of ambiguous lesion regions was a key outcome.

Conclusions:

• ALA-Net effectively integrates contextual and spatial information for robust 3D medical image segmentation.
• The proposed network shows excellent generalization capabilities for various 3D medical image segmentation tasks.
• ALA-Net offers significant benefits for clinical applications requiring precise tumor segmentation.