Adaptive Feature Medical Segmentation Network: an adaptable deep learning paradigm for high-performance 3D brain...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurological Diagnostics

Background:

• Accurate brain lesion detection and segmentation are critical in neurological diagnostics.
• Traditional methods struggle with complex lesion morphology, often sacrificing detail for efficiency or vice versa.
• Balancing performance and precision in compute-intensive medical imaging is a key research challenge.

Purpose of the Study:

• Introduce a novel, lightweight encoder-decoder network for brain lesion segmentation.
• Address the trade-off between computational efficiency and segmentation accuracy.
• Provide a scalable framework for advanced medical image processing.

Main Methods:

• Developed the Adaptive Feature Medical Segmentation Network (AFMS-Net) with two encoder variants: Single Adaptive Encoder Block (SAEB) and Dual Adaptive Encoder Block (DAEB).
• SAEB uses squeeze-and-excite for efficient segmentation of key areas.
• DAEB employs channel spatial attention for fine-grained, multi-class delineation.
• Integrated a Segmentation Path (SegPath) module for enhanced feature extraction and stability.

Main Results:

• AFMS-Net demonstrated exceptional performance on benchmark datasets (BRATs 2021, ATLAS 2021, ISLES 2022).
• The network achieves high precision in segmenting complex brain lesions.
• The lightweight architecture effectively handles demanding segmentation tasks.

Conclusions:

• AFMS-Net successfully balances performance and computational efficiency in brain lesion segmentation.
• The adaptive encoder variants allow tailoring the network to specific speed and feature requirements.
• This work advances state-of-the-art lesion segmentation and offers a foundation for future medical imaging research.