AI models for MRI segmentation show limited generalisability despite intensity normalisation

Researchers conducted a systematic benchmark of seven intensity normalization methods for 3D knee MRI segmentation using a 3D U-Net model. The study found that while methods like Z-score, Nyúl histogram matching, and CLAHE demonstrated greater robustness on external datasets, the overall impact of intensity normalization was limited compared to the significant performance drop caused by domain shift. This suggests that while normalization is a factor, complementary strategies are crucial for robust clinical deployment of deep learning models in medical imaging. AI

IMPACT Highlights the limitations of current AI generalisation techniques in medical imaging, suggesting a need for new strategies beyond intensity normalisation.

RANK_REASON The cluster contains an academic paper detailing a systematic benchmark of methods for medical imaging analysis. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• 3D U-Net
• arXiv
• Contrast Limited Adaptive Histogram Equalization
• Gaussian Mixture Model
• IWOAI 2019
• Nyúl histogram matching

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →