NeighborNet: Learning Intra- and Inter-Image Pixel Neighbor Representation for Breast Lesion Segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Accurate breast lesion segmentation in ultrasound images is crucial for computer-aided breast cancer diagnosis.
• Existing methods combining CNNs and attention struggle with blurry boundaries and irregular shapes due to independent feature extraction.
• Current transformer-based methods have limited global information capture within single images, ignoring cross-image semantic relations.

Purpose of the Study:

• To develop a flexible method for integrating global and local context within and across images for improved breast lesion segmentation.
• To address the limitations of independent feature extraction and limited cross-image information in existing models.
• To enhance the modeling of lesion morphology and boundaries, particularly in cases of irregular shapes and blurry edges.

Main Methods:

• Proposed NeighborNet, a pixel neighbor representation learning method for adaptive context representation.
• Introduced two neighbor layers to investigate neighbor number and distribution properties.
• Enabled NeighborNet to adaptively evolve into transformer or CNN structures based on pixel-level feature requirements.

Main Results:

• NeighborNet demonstrated state-of-the-art performance on three diverse ultrasound datasets.
• The method effectively handles irregular lesion morphologies and blurry boundaries.
• Flexible integration of global and local context improved segmentation accuracy.

Conclusions:

• NeighborNet offers a robust and adaptive approach for breast lesion segmentation in ultrasound imaging.
• The proposed method overcomes limitations of previous techniques by flexibly integrating multi-level contextual information.
• NeighborNet shows significant potential for advancing computer-aided breast cancer diagnosis.