Multi-level spatio-relational segformer (MLSRS-SegFormer): A novel vision transformer with adaptive spatial induction...

Area of Science:

• Computer Vision
• Medical Image Analysis
• Artificial Intelligence

Background:

• Medical image segmentation is crucial for precision medicine.
• Current Vision Transformers (ViTs) face a trade-off between global context and local boundary detail, particularly with high-variance clinical data.

Purpose of the Study:

• To introduce a novel Vision Transformer architecture, the Multi-Level Spatio-Relational SegFormer (MLSRS-SegFormer), to overcome the limitations of existing models in medical image segmentation.
• To enhance semantic segmentation accuracy and boundary delineation for clinical applications.

Main Methods:

• Developed the MLSRS-SegFormer, incorporating Adaptive Patch Weighting for dynamic feature induction, Hausdorff-bias Attention for spatial prioritization, and Relative Positional Encoding (RPE) for adaptive spatial understanding.
• Evaluated the model through comparative experiments and Bland-Altman analyses.

Main Results:

• MLSRS-SegFormer achieved state-of-the-art performance with the highest mean Intersection over Union (mIoU) of 0.968 and mean Dice Similarity Coefficient (mDSC) of 0.980.
• Demonstrated exceptional boundary precision with the lowest Hausdorff distance 95th percentile (HD95) of 1.1668.
• Bland-Altman analyses confirmed minimal systematic bias and high consistency in segmentation, vital for clinical use.

Conclusions:

• The MLSRS-SegFormer architecture offers significant improvements in medical image segmentation, particularly in boundary accuracy.
• Its robust performance and precision make it highly suitable for critical clinical applications, despite a slightly longer inference time.