SAN-Net: Learning generalization to unseen sites for stroke lesion segmentation with self-adaptive normalization

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Automatic stroke lesion segmentation on MRI is crucial for diagnosing cerebrovascular diseases.
• Deep learning models struggle with generalization due to inter-site data discrepancies and lesion variations.

Purpose of the Study:

• To introduce a self-adaptive normalization network (SAN-Net) for improved generalization in stroke lesion segmentation.
• To address challenges posed by variations in scanners, protocols, populations, and lesion characteristics.

Main Methods:

• Developed Masked Adaptive Instance Normalization (MAIN) to minimize inter-site discrepancies by learning affine parameters.
• Utilized a gradient reversal layer with a site classifier to promote site-invariant feature learning in the U-net encoder.
• Introduced Symmetry-Inspired Data Augmentation (SIDA) to increase sample size and reduce memory usage.

Main Results:

• SAN-Net demonstrated superior performance on the ATLAS v1.2 dataset under a leave-one-site-out cross-validation setting.
• Quantitative metrics and qualitative comparisons showed significant improvements over recently published methods.
• The network effectively standardized MR images into a site-unrelated style.

Conclusions:

• SAN-Net achieves adaptive generalization for stroke lesion segmentation on unseen sites.
• The combination of MAIN, gradient reversal, and SIDA enhances model robustness and efficiency.
• This approach offers a promising solution for reliable automatic stroke lesion segmentation in diverse clinical settings.