💻 computer science

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

This study systematically evaluates seven intensity normalization methods for 3D knee MRI meniscus segmentation, finding that while techniques like Z-score and histogram matching improve cross-domain generalizability, their impact remains limited compared to the significant performance drop caused by domain shifts between datasets.

Original authors: Oliver Mills, Philip Conaghan, Samuel Relton

Published 2026-07-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Oliver Mills, Philip Conaghan, Samuel Relton

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize a specific type of cookie in a bakery. You show the robot thousands of pictures of chocolate chip cookies, but there's a catch: every time you take a photo, the lighting changes. Sometimes the lights are bright and yellow, sometimes they are dim and blue, and sometimes the camera lens is a little dirty. If you only teach the robot under perfect, bright lights, it might get confused when it sees a cookie in a dim corner. This is a huge problem in the world of medical imaging, specifically with Magnetic Resonance Imaging (MRI). MRI machines are like those tricky cameras; even if they are scanning the same body part, the "brightness" or intensity of the images can vary wildly depending on the machine, the settings, or the patient. Doctors and scientists use deep learning (a type of AI) to help spot problems in these images, but if the AI is too sensitive to these lighting changes, it might fail when it moves from one hospital to another. To fix this, researchers use "normalization," which is like a photo-editing tool that tries to make all the images look like they were taken under the same perfect lighting before the AI even sees them. The big question is: which photo-editing tool actually works best to make the AI a master cookie-recognizer, no matter where it is?

In this study, a team of researchers set out to find the best "photo-editing" tool for a very specific and tricky task: teaching an AI to find the meniscus (a small, shock-absorbing cartilage) in knee MRI scans. They didn't just guess; they ran a systematic race, testing seven different normalization methods to see which one helped the AI perform best. They trained their AI model on a dataset called IWOAI 2019, which contained knee scans from one group of patients, and then they put the model to the ultimate test: seeing if it could still work on a completely different set of knee scans from a different hospital (the SKM-TEA dataset). This is like training a robot in one bakery and then sending it to a totally different bakery across town to see if it can still find the cookies.

The results were a mix of "good news" and "reality check." When the AI was tested on images from the same hospital where it was trained, all seven methods performed almost identically. It was like a tie; the specific photo-editing tool didn't matter much when the lighting was familiar. However, when the AI faced the new, different hospital data, the differences started to show up. The study found that three methods stood out as slightly more robust: Z-score (a standard way of adjusting brightness), Nyúl histogram matching (a technique that tries to match the "shape" of the brightness distribution to a template), and CLAHE (a method that boosts contrast in small areas). Among these, Nyúl matching and Z-score were the top contenders, with Nyúl showing the best ability to keep the volume measurements accurate.

But here is the most important twist in the story: while these differences were real and statistically significant, they were actually quite small. The researchers found that the drop in performance when moving from the training hospital to the new hospital was massive—about a 10% drop in accuracy. In contrast, the difference between the best normalization method and the worst was only about 1%. It's like realizing that while using a specific brand of lens cleaner helps your camera take slightly sharper photos, the real problem is that you are trying to take a photo in a pitch-black room. The study suggests that while picking the right normalization method (like Nyúl or Z-score) does help the AI generalize to new data, it is not a magic bullet. The biggest hurdle remains the "domain shift"—the fundamental differences between how different hospitals scan knees. The authors conclude that while these simple normalization tricks are useful, they aren't enough on their own to solve the problem of making AI work perfectly everywhere; we will likely need other, more advanced strategies to truly bridge the gap between different medical centers.

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

Problem Statement
The clinical deployment of deep learning models in medical imaging is hindered by "domain shift," where models trained on data from specific scanners or protocols fail to generalize to unseen institutions. While deep learning has shown promise in medical image segmentation, performance often degrades significantly when applied to external datasets due to variations in scanner types, imaging protocols, and patient populations. In Magnetic Resonance Imaging (MRI), signal intensities lack intrinsic meaning and vary widely, making standardisation crucial. Although advanced domain adaptation techniques exist, they often require access to target-domain data or additional training, which is impractical in many deployment scenarios. Consequently, there is a need to evaluate whether standard, "out-of-the-box" intensity normalisation methods—requiring no retraining—can improve model robustness against domain shift, particularly in musculoskeletal imaging where such studies are limited compared to brain or breast MRI.

Methodology
This study systematically compared seven intensity normalisation methods for 3D meniscus segmentation using a 3D U-Net architecture (specifically the nnU-Net framework).

Datasets:
- Training/Internal Test: The IWOAI 2019 dataset (subset of the Osteoarthritis Initiative), containing 176 3D sagittal DESS MRI scans from 88 patients.
- External Test: The SKM-TEA dataset, containing 155 3D quantitative DESS (qDESS) MRI scans from patients with knee pathology, acquired on different scanners with different protocols.
Preprocessing: Images were cropped to focus on the knee region. For the external dataset, the dual-echo qDESS images were combined using a root-sum-of-squares method.
Normalisation Methods Evaluated:
1. Z-score: Standardisation using image-specific mean and standard deviation.
2. Min-max scaling: Linear scaling to [0, 1].
3. Robust Min-max: Scaling after clipping intensities to the 1st and 99th percentiles.
4. Histogram Equalisation (HE): Global equalisation of the intensity histogram.
5. Contrast-Limited Adaptive Histogram Equalisation (CLAHE): Local equalisation with a clip limit to prevent noise amplification.
6. Nyúl Standardisation: A histogram matching method using landmarks derived from a training template.
7. Gaussian Mixture Model (GMM): A method adapted from brain MRI "white stripe" normalisation, fitting a GMM to the intensity distribution and standardising based on a specific component peak.
Experimental Design: Models were trained using 5-fold cross-validation on the IWOAI 2019 data. Performance was evaluated on both the internal test split and the external SKM-TEA dataset.
Metrics: Volumetric agreement (Dice Similarity Coefficient - DSC, Lin's Concordance Correlation Coefficient - CCC, Mean Absolute Error - MAE) and spatial accuracy (95th percentile Hausdorff Distance - HD95). Statistical significance was assessed using Linear Mixed-Effects Models (LMM) to account for repeated measures across folds and images.

Key Results

Internal Performance: On the internal test set (IWOAI 2019), performance differences between methods were minimal. Z-score, Min-max, and GMM achieved the highest mean DSC (~89.05%), while HE performed significantly worse. No significant differences were found for HD95.
External Performance: On the external test set (SKM-TEA), a significant performance drop was observed across all methods (mean DSC dropped from ~89% to ~78-79%), highlighting the severity of domain shift.
- Top Performers: Nyúl standardisation achieved the highest DSC (79.10%), followed closely by Z-score (78.95%) and CLAHE (78.86%).
- Statistical Significance: LMM analysis confirmed that Nyúl, Z-score, and CLAHE significantly outperformed Min-max, Robust Min-max, HE, and GMM on the external dataset.
- Volumetric Bias: Nyúl demonstrated the best volumetric consistency with ground truth, whereas Z-score tended to slightly underestimate meniscus volume.
- Magnitude of Effect: While statistically significant, the performance gap between the best and worst normalisation methods on external data was small (~~1% DSC). In contrast, the performance drop caused by the domain shift itself was substantial (~~10% DSC). The marginal $R^2$ indicated that the choice of normalisation method explained only 0.5% of the total variance in DSC scores.

Significance and Claims
The paper concludes that while intensity normalisation has a measurable and statistically significant effect on model generalisability, its impact is limited relative to the effects of domain shift.

Robustness: Methods such as Nyúl histogram matching, Z-score, and CLAHE demonstrated greater robustness to cross-dataset variations compared to standard scaling or GMM-based approaches.
Practical Implication: The study highlights that simple, preprocessing-only normalisation strategies can provide a practical baseline for improving out-of-distribution performance without requiring target-domain data or complex retraining.
Limitations of Normalisation: The authors emphasize that normalisation alone cannot overcome the significant performance degradation caused by domain shift (differences in protocols, scanners, and patient populations). Therefore, robust clinical deployment will likely require complementary strategies beyond standard intensity normalisation.
Domain Specificity: The findings suggest that methods effective in brain MRI (like GMM-based white stripe normalisation) may not translate directly to musculoskeletal imaging due to differences in intensity profiles and tissue characteristics.

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

More like this