Source-linked AI summary
Evaluating the Impact of Intensity Normalization on MR Image Synthesis
Jacob C. Reinhold, Blake E. Dewey, Aaron Carass, Jerry L. Prince
TL;DR
MR image synthesis requires intensity transformations between contrasts, yet the optimal input scaling is unknown and MR images lack a standard intensity scale. The paper evaluates seven normalization algorithms with three synthesis methods and finds that preprocessing normalization improves synthesis, especially for deep neural networks. The authors also report that synthesis is robust to the normalization algorithm, while unnormalized inputs can produce failures and slice-to-slice intensity inconsistencies.
Problem
MR image synthesis depends on comparable intensity properties across training and testing data, but MR images lack a standard intensity scale and the optimal input scaling is unknown.
Method
The paper compares seven intensity-normalization algorithms and three synthesis routines for synthesizing T2-w and FLAIR images from T1-w brain MRI.
Results
Synthesis methods are substantially improved by intensity-normalization preprocessing, especially DNN synthesis, while no normalization method consistently differs significantly across metrics or synthesis algorithms.
Takeaways & Limitations
Normalization is suggested before training or testing patch-based DNNs because DNN performance is better across all metrics on normalized images.
Takeaways & Limitations
Slice-to-slice intensity variation remains partly attributable to the paper’s use of a 2D synthesis method.
Abstract
from arXiv · showhide
Image synthesis learns a transformation from the intensity features of an input image to yield a different tissue contrast of the output image. This process has been shown to have application in many medical image analysis tasks including imputation, registration, and segmentation. To carry out synthesis, the intensities of the input images are typically scaled--i.e., normalized--both in training to learn the transformation and in testing when applying the transformation, but it is not presently known what type of input scaling is optimal. In this paper, we consider seven different intensity normalization algorithms and three different synthesis methods to evaluate the impact of normalization. Our experiments demonstrate that intensity normalization as a preprocessing step improves the synthesis results across all investigated synthesis algorithms. Furthermore, we show evidence that suggests intensity normalization is vital for successful deep learning-based MR image synthesis.
1. INTRODUCTION
MR image synthesis learns intensity transformations between tissue contrasts, but its accuracy depends on a shared intensity scale that MR images lack. This paper evaluates seven normalization methods with three synthesis algorithms to assess normalization’s impact.
- MR synthesis can generate unavailable tissue contrasts for applications including image imputation, registration, and segmentation.
- Accurate synthesis depends heavily on a standard intensity scale across training images and similar intensity properties between training and testing data.For example, the mean intensity of white matter should be consistent across input images.
- MR synthesis presents this scaling problem because MR images do not have a standard intensity scale.
- The study evaluates seven intensity-normalization methods on T1-w, T2-w, and FLAIR brain images, comparing synthesis of T2-w and FLAIR from T1-w using three algorithms.
2. METHODS
The methods comprise seven intensity-normalization algorithms and three synthesis routines. Together, they define the experimental comparison of normalization choices across synthesis approaches.
- The study considers seven normalization algorithms: Z-score, FCM-based, GMM-based, KDE-based, histogram matching, WhiteStripe, and RAVEL.
- The three synthesis routines are polynomial regression, random forest regression, and deep-neural-network-based synthesis.
- The methods formalize each MR image over three-dimensional voxel coordinates together with a corresponding brain mask.The brain mask identifies indices corresponding to the brain location in the image.
2.1 Normalization
The paper reviews seven intensity normalization algorithms for MR brain images, ranging from statistical scaling and tissue-based methods to histogram matching and technical-variation removal.
- Z-score: Z-score normalization uses the brain-mask mean and standard deviation to scale each image's intensities.
- Tissue-based normalization: FCM- and GMM-based methods normalize the image using a white-matter intensity estimate obtained from fuzzy clustering or a three-component mixture model.
- Kernel Density Estimate-based: KDE-based normalization smooths the brain-mask intensity histogram, identifies the white-matter peak, and uses that peak to normalize the image.
- Piecewise Linear Histogram Matching: Piecewise linear histogram matching learns a standard histogram from image-set landmarks and maps each image to it along percentile-defined segments.
- WhiteStripe: WhiteStripe performs Z-score normalization using a 10% intensity segment around the normal-appearing white-matter peak.
- RAVEL: RAVEL extends WhiteStripe by modeling CSF-associated technical variation, estimating its coefficients through singular-vector bases and voxel-wise regression, and removing it from images.
2.2 Synthesis
The paper evaluates polynomial regression, random forest regression, and deep neural network methods for synthesizing one MR contrast from another. Each method learns mappings from source-image intensities or patches to target-image intensities.
- Image synthesis is formulated as learning a parametric or non-parametric mapping between intensity distributions of different contrasts.
- Polynomial regression: Polynomial regression uses seven-voxel source patches and corresponding target center voxels sampled from 100,000 brain-mask locations.Across M images, this produces an (M · 100,000) × 7 source feature matrix and an (M · 100,000) × 1 target feature matrix.
- Random forest regression: Random forest regression uses 25-voxel source patches spanning local neighbors and primary directions at distances of 3, 5, and 7 voxels.
- Deep neural network synthesis: The deep neural network method uses a four-level U-net trained on 128 × 128 patches from axial, sagittal, and coronal orientations.Instance normalization and leaky ReLUs are used because the normalized images can contain negative values.
2.3 Quality Assessment
Synthesis quality is evaluated against held-out ground-truth images using three metrics selected to remain comparable across normalization ranges. The metrics are normalized cross-correlation, mean structural similarity, and mutual information.
- Quality metrics compare synthesized results with ground-truth images that were not used to train the synthesis methods.
- The evaluation uses normalized cross-correlation, mean structural similarity, and mutual information.
- MSE and PSNR are excluded because differing normalization routines scale the data to different ranges, making those metrics difficult to compare.
3. RESULTS
The experiments use 18 healthy-subject data sets split evenly into training and testing, with quality summarized across normalization schemes and synthesis algorithms. Unnormalized synthesis produced failure cases, whereas FCM-normalized synthesis was successful in the illustrated T1-to-FLAIR and T1-to-T2 examples.
- Experimental setup: Evaluation uses 18 healthy-subject data sets, with nine assigned to training and nine to testing.T1-w, T2-w, and FLAIR images are resampled, bias-field corrected, registered, and brain-masked before normalization and synthesis.
- Experimental setup: Bar charts report mean quality metrics and bootstrapped 95% confidence intervals across testing data sets for each normalization scheme and synthesis algorithm.Wilcoxon signed-rank tests compare normalized methods against corresponding unnormalized results.
- Normalization comparison: Synthesis results are robust to the choice of normalization algorithm, remaining stable around similar levels across metrics.No normalization algorithm consistently reached the Bonferroni-corrected significance threshold for any metric and synthesis algorithm in either task.
- Synthesis results: Unnormalized T1-to-FLAIR and T1-to-T2 synthesis includes failure cases, while the corresponding FCM-normalized examples are successfully synthesized.
- T1-to-FLAIR results: Figure 1 compares Raw, ZS, and WS T1-to-FLAIR quality metrics; asterisks denote significance versus Raw at p < 0.05, with 95% confidence-interval error bars.
4. DISCUSSION AND CONCLUSION
Intensity normalization substantially improves synthesis, especially for DNNs, while performance is robust across the normalization methods evaluated. Unnormalized inputs can produce failures and slice-to-slice intensity inconsistencies, even under controlled acquisition conditions.
- Synthesis methods are substantially improved by intensity normalization, especially DNN synthesis, with no statistically significant difference among the normalization methods presented.
- Unnormalized DNN synthesis fails for particular T1-w inputs whose intensity histograms differ from most training images.The problematic histogram compressed the grey-matter peak near the training-set white-matter peak average.
- Unnormalized synthesis can appear reasonable in the synthesis plane while showing substantial neighboring-slice intensity variation in another plane.This slice-to-slice variation is not observed in normalized synthesis results.
- Unnormalized synthesis failed even when training and testing images came from the same cohort, scanner, and pulse sequence and involved healthy patients.
- Because DNN performance is better across all metrics with normalized images, normalization is suggested before training or testing patch-based DNNs.