Attention based multi-scale nested network for biomedical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) have advanced biomedical image segmentation.
• Medical image segmentation faces challenges due to intra-sample variability and inter-sample specificity.
• Existing models often overlook the unique challenges of medical image segmentation.

Purpose of the Study:

• To propose a novel architecture, Attention based multi-scale nested network (AMNNet), for efficient biomedical image segmentation.
• To address the limitations of current models in handling intra-sample variability in medical datasets.
• To improve the accuracy and efficiency of medical image segmentation.

Main Methods:

• Developed AMNNet, incorporating ReSidual U-CBAM (RSUC) modules, convolutional stages, a MLP stage, and Convolutional Block Attention Modules (CBAM).
• Introduced a lightweight CBAM to focus on relevant regions and suppress irrelevant features.
• Designed RSUC modules to capture multi-scale contextual information by combining different receptive fields.

Main Results:

• AMNNet demonstrated superior performance compared to existing methods across multiple benchmark datasets (ISIC2018, CVC-ClinicDB, CVC-ColonDB, BUSI, GlaS).
• Achieved high Dice Similarity Coefficients (DSC): 91.35% (ISIC2018), 90.01% (CVC-ClinicDB), 90.80% (CVC-ColonDB), 81.61% (BUSI), and 94.31% (GlaS).
• The proposed architecture effectively handles the specificity and variability inherent in medical images.

Conclusions:

• AMNNet offers an effective solution for challenging biomedical image segmentation tasks.
• The integration of attention mechanisms and multi-scale processing enhances segmentation accuracy.
• AMNNet provides a promising advancement for automated medical image analysis.