Explainable multi-module semantic guided attention based network for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated medical image segmentation is vital for diagnostics and treatment planning.
• Convolutional Neural Networks (CNNs) have advanced segmentation but struggle with variations in object size, shape, and position.
• Current CNNs lack explainability, hindering clinical adoption.

Purpose of the Study:

• To develop an explainable and highly accurate medical image segmentation network.
• To address limitations of existing CNNs regarding segmentation variability and clinical interpretability.
• To integrate multi-scale and semantic guided attention mechanisms for improved feature extraction.

Main Methods:

• Proposed the explainable multi-module semantic guided attention based network (MSGA-Net).
• Introduced a multi-scale attention module (MSA) for salient feature extraction at various scales.
• Developed a semantic region-guided attention mechanism (SRGA) with location, channel-wise, and edge attention modules.
• Employed a fine-tuning strategy with SRGA to enhance region significance and reduce noise.

Main Results:

• MSGA-Net demonstrated substantial performance improvements across all metrics on dermoscopic, CT, and MRI datasets.
• The model achieved higher accuracy and explainability compared to existing state-of-the-art models.
• Attention feature maps provided enhanced interpretability for clinical insights.

Conclusions:

• MSGA-Net offers a significant advancement in explainable and accurate medical image segmentation.
• The proposed attention mechanisms effectively handle variations in medical image segmentation tasks.
• MSGA-Net shows promise for clinical applications requiring reliable and interpretable segmentation results.