A Systematic Benchmark of Intensity Normalisation Methods for 3D Knee MRI Segmentation and Cross-Domain Generalisability

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

A systematic benchmark compared seven intensity normalisation methods for 3D U-Net meniscus segmentation from knee MRI, evaluating their impact on model generalisability. Researchers from the University of Leeds trained models on the IWOAI 2019 dataset and tested them on both internal and external data (SKM-TEA). The methods included Z-score, min-max scaling, robust min-max scaling, histogram equalisation (HE), contrast-limited adaptive histogram equalisation (CLAHE), Nyúl standardisation, and Gaussian Mixture Model (GMM) normalisation. While internal performance differences were minimal, external evaluation on SKM-TEA revealed small but statistically significant advantages for Nyúl, Z-score, and CLAHE, which outperformed other methods by approximately 1% in Dice Similarity Coefficient (DSC). However, all methods experienced a substantial performance drop of about 10% DSC between internal and external datasets, indicating that domain shift has a much greater impact than the choice of normalisation method. The study highlights the limited role of simple intensity normalisation in mitigating domain shift for robust clinical deployment.

Key takeaway

For Machine Learning Engineers deploying 3D knee MRI segmentation models across different clinical sites, selecting an intensity normalisation method like Nyúl, Z-score, or CLAHE can offer a small, statistically significant improvement in cross-domain generalisability. However, you should recognize that these methods only account for about 0.5% of performance variance. Focus your efforts on more comprehensive domain adaptation strategies, as domain shift between datasets causes a much larger performance drop (around 10% DSC) than normalisation choice.

Key insights

Intensity normalisation minimally improves 3D knee MRI segmentation generalisability; domain shift remains the primary performance challenge.

Principles

• Internal performance differences are negligible.
• External data reveals small, significant method differences.
• Domain shift impact far outweighs normalisation choice.

Method

Seven intensity normalisation methods (Z-score, min-max, robust min-max, HE, CLAHE, Nyúl, GMM) were systematically compared using a 3D U-Net for meniscus segmentation on IWOAI 2019 and SKM-TEA datasets.

In practice

• Prioritize Nyúl, Z-score, or CLAHE for cross-domain tasks.
• Expect ~1% DSC gain from optimal normalisation.
• Address domain shift with complementary strategies.

Topics

• 3D MRI Segmentation
• Intensity Normalisation
• Domain Shift
• Medical Imaging
• Knee Meniscus
• Deep Learning Models

Code references

• denizlab/2019_IWOAI_Challenge
• StanfordMIMI/skm-tea
• oliverjm1/mri_normalisation

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Systematic Benchmark of Intensity Normalisation Methods for 3D Knee MRI Segmentation and Cross-Domain Generalisability — Intensity normalization improves MRI segmentation generalizability, but domain shift remains a larger challenge for clinical deployment.
• Domain Shift in Echocardiography: Interpretable Quantification and Prediction of Cross-Dataset Left Ventricular Segmentation — Echocardiographic domain shift is measurable and its impact on LV segmentation can be predicted and mitigated.
• CrossPan: A Comprehensive Benchmark for Cross-Sequence Pancreas MRI Segmentation and Generalization — Cross-sequence domain shift is the primary barrier to robust pancreas MRI segmentation.
• Lesion Segmentation in Moderate to Severe Traumatic Brain Injury: An nnU-Net Based Approach with Adaptive Normalization in the AIMS-TBI 2025 Challenge — Anatomically constrained adaptive normalization within nnU-Net significantly improves msTBI lesion segmentation performance.
• ConRad: Efficient Conformal Prediction for Radiomics — ConRad improves radiomic feature reliability by using adaptive conformal prediction with segmentation uncertainty.
• Contrast-Informed Augmentation and Domain-Adversarial Training for Adult-to-Neonatal MR Reconstruction Generalization — Contrast-informed augmentation and domain-adversarial training enhance deep learning MR reconstruction generalization from adult to neonatal data.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.