Adaptive multi-stage domain unlearning for white-matter lesion segmentation

Area of Science:

• Medical imaging analysis
• Machine learning for medical applications
• Neuroimaging

Background:

• Inter-scanner variability in MRI degrades scan quality and diagnostic accuracy.
• Domain shift from unseen scanners poses challenges for AI model generalization.
• Existing domain adaptation/generalization methods' efficacy depends on training supervision levels.

Purpose of the Study:

• To develop an adaptive multi-stage domain unlearning (ADMU) technique for robust MRI segmentation across unseen scanner domains.
• To enhance the generalization capabilities of deep learning models in medical imaging.
• To improve white-matter lesion segmentation performance in the presence of inter-scanner variability.

Main Methods:

• Proposed an adaptive multi-stage domain unlearning (ADMU) technique integrated into the nnU-Net framework.
• Employed deep supervision with sequential domain classifier unlearning at deep encoder stages to reduce domain-discriminative features.
• Implemented an adaptive backpropagation schedule for unlearning, balancing task impact.

Main Results:

• ADMU demonstrated consistent, robust, and improved cross-dataset segmentation performance compared to baseline nnU-Net variants.
• Achieved enhanced lesion sensitivity with balanced false detections, leading to superior segmentation quality (overlap, volume error).
• Adaptive scheduling of unlearning mitigated adverse impacts on the primary segmentation task; intensity preprocessing was detrimental.

Conclusions:

• The ADMU strategy is effective and complementary to data augmentation for robust white-matter lesion segmentation.
• The method simplifies preprocessing, relying solely on FLAIR modality and spatial normalization, achieving optimal cross-dataset performance.
• The proposed ADMU technique offers a promising solution for domain generalization in medical image analysis.