SwiftMSeg: lightweight multi-scale local-global context modeling with transformer for medical image segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning architectures

Background:

• Accurate medical image segmentation demands simultaneous fine boundary localization and contextual understanding.
• Achieving both is challenging, especially in lightweight deep learning models.
• Existing methods often struggle to balance precision and efficiency.

Purpose of the Study:

• To introduce SwiftMSeg, a novel lightweight encoder-decoder framework for medical image segmentation.
• To address the boundary-context challenge in medical image analysis.
• To develop a computationally efficient yet accurate segmentation model.

Main Methods:

• Developed SwiftMSeg, integrating a convolutional encoder, a transformer-based local-global-local module, and a hierarchical multi-scale decoder.
• Combined progressive multi-scale refinement with global context modeling for improved segmentation.
• Evaluated on colonoscopy, pathology, ultrasound, and MRI datasets.

Main Results:

• SwiftMSeg achieved high Dice scores across modalities (e.g., 0.896 for colonoscopy, 0.870 for MRI).
• Demonstrated improved boundary localization (lower Hausdorff distance) and stable segmentation.
• Showcased significant domain-independent generalization capabilities on an external dataset.
• Achieved high efficiency with only 4.48M parameters and 0.940 GFLOPs, a ~53x reduction in computational cost compared to U-Net baselines.

Conclusions:

• SwiftMSeg effectively balances boundary localization and contextual understanding in medical image segmentation.
• The framework offers a practical, scalable, and computationally efficient solution for diverse medical imaging applications.
• SwiftMSeg outperforms existing methods in accuracy and efficiency, making it suitable for real-world deployment.