Lightweight Medical Image Segmentation with UNet Architecture

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence in dermatology

Background:

• Accurate segmentation of skin lesions is crucial for diagnosis and treatment.
• Existing segmentation models often face challenges with computational cost and boundary delineation.

Purpose of the Study:

• To propose two novel, efficient, and lightweight semantic segmentation models (MCS-Net and DAC-Net) for skin lesion segmentation.
• To enhance feature representation and boundary delineation in medical images using attention mechanisms and multi-scale contextual cues.

Main Methods:

• Development of a symmetric six-level U-shaped architecture.
• Integration of attention mechanisms and multi-scale contextual information within the UNet framework.
• Evaluation using standard segmentation metrics such as Dice Similarity Coefficient (DSC) and mean Intersection over Union (mIoU).

Main Results:

• The proposed MCS-Net and DAC-Net models demonstrated superior performance in skin lesion segmentation.
• Significant improvements in DSC and mIoU were observed compared to existing methods.
• The models achieved enhanced feature learning and more precise boundary delineation.

Conclusions:

• The developed lightweight segmentation models offer an effective solution for skin lesion analysis.
• The integration of attention and multi-scale context is a promising approach for improving medical image segmentation.
• The proposed architectures are efficient and suitable for practical applications requiring high accuracy.