Intensity warping for multisite MRI harmonization

Julia Wrobel, Ml Martin, Rohit Bakshi, Peter A. Calabresi, Mark Elliot, D Raolf, Rc Gur, Re Gur, Rg Henry, Govind Nair, Jiwon Oh, Nico Papinutto, Daniel Pelletier, Ds Reich, William D. Rooney, T. E. Satterthwaite, W. Stern, Karthik Prabhakaran, Nancy L. Sicotte, Rt Shinohara, Jeff Goldsmith, on behalf of the Naims CooperativebioRxiv 2019 · DOI

TL;DR

This paper introduces "mica," a novel method for harmonizing MRI scans from different sites and scanners. By aligning the intensity distributions of scans from the same individual, mica effectively reduces technical variability while preserving biological differences, making multisite neuroimaging studies more reliable.

What Problem Does This Paper Address?

Multisite neuroimaging studies are crucial for large-scale research, but they face a significant challenge: "scanner effects." These are technical variations introduced by differences in MRI scanners, software, or acquisition protocols across various locations. These effects can obscure true biological signals, lead to misleading results, and make it difficult to compare data collected at different sites. Existing methods often struggle to fully address these variations, especially when they are complex and non-linear, or they may inadvertently remove biologically relevant information. This paper aims to develop and validate a method that specifically targets and reduces these scanner effects without compromising the integrity of the underlying biological data.

Methods

The core of the proposed method, "mica" (multisite image harmonization by CDF alignment), lies in its innovative approach to aligning the intensity distributions of MRI scans. Instead of directly manipulating voxel intensities, mica focuses on the cumulative distribution functions (CDFs) of these intensities.

Here's a breakdown of the key steps:

  1. CDF Estimation: For each scan, the probability density function (PDF) and its corresponding CDF are calculated.
  2. Alignment via Warping: The method estimates a non-linear, monotonic transformation (a "warping function") for the intensity values of one scan. This function is designed so that the transformed scan's intensity CDF perfectly matches the CDF of a chosen "target" scan. This alignment is achieved using standard curve registration techniques.
  3. Harmonized Image Generation: The estimated warping function is then applied to the original voxel intensities of the scan, producing a harmonized image.

This approach allows for the identification and removal of scanner effects by leveraging multiple scans from the same subject, effectively creating a subject-specific template for harmonization. The method is designed to be applied to the entire brain, avoiding the need for spatial registration to a common template, which can sometimes reduce image resolution.

Datasets

The authors evaluated their method using two distinct datasets:

  1. NAIMS pilot study: This dataset involved a single subject with multiple sclerosis (MS) who underwent scan-rescan imaging at seven different sites across the US. This dataset was used to demonstrate how mica can reduce technical variability across sites while preserving the ability to detect MS lesions.
  2. trio2prisma study: This dataset consisted of ten healthy subjects who were scanned on two different MRI machines (a Siemens Trio and a Prisma) within a short timeframe. This allowed the researchers to study systematic, non-linear differences between scanners and to assess mica's ability to create mappings between scanners.

Key Insights and Ideas

  • Distinguishing Intensity Unit Effects from Scanner Effects: The paper clearly differentiates between "intensity unit effects" (arbitrary intensity values within a scanner) and "scanner effects" (technical artifacts across scanners/sites). This distinction is crucial for developing targeted harmonization strategies.
  • Leveraging Within-Subject Variability: The core innovation of mica is its ability to use multiple scans from the same individual to understand and correct for scanner effects. This is a powerful approach because it assumes that differences between these scans are primarily technical, not biological.
  • CDF Alignment as a Robust Harmonization Strategy: Focusing on CDFs, rather than raw intensities, provides a more robust way to align distributions, especially when dealing with complex, non-linear scanner effects.
  • Preserving Biological Variability: A key finding is that mica successfully reduces technical variability while preserving biological differences across subjects, a critical requirement for any harmonization method.

Key Results

  • Reduced Site-to-Site Variability in Lesion Volume: In the NAIMS study, mica harmonization, when paired with White Stripe normalization, significantly reduced the variance in T2 lesion volume measurements across sites compared to White Stripe alone (variance of 11.8 ml² vs. 17.1 ml²).
  • Effective Harmonization in the trio2prisma Study: For the trio2prisma data, mica effectively aligned the intensity distributions of scans from the Trio and Prisma scanners, reducing within-subject technical variability.
  • Preservation of Inter-Subject Variability: Unlike histogram matching, which tended to eliminate biological differences, mica-harmonized scans showed similar across-subject variability to the original target scans (Prisma scans), indicating preservation of biological information.
  • Qualitative Improvement: Visual inspection of images showed that mica harmonization brightened the contrast between white and gray matter without distorting biological features, whereas histogram matching drastically altered image appearance.

Limitations and Future Work

  • Imperfect Average Scanner Variability: The authors acknowledge that their method, by using a subset of patients to estimate average scanner variability, is imperfect. This average may not perfectly capture the true variability for every subject.
  • Potential for Covariate Interaction: Evidence suggests that scanner effects can vary with covariates like gender and age. Future work could extend mica to incorporate such covariates to further refine harmonization.
  • Generalizability: While the method is presented as generalizable to the entire brain, further validation across different imaging modalities and a wider range of scanner types and sites would be beneficial.
  • Software Availability: The authors have released an R package for mica, encouraging its adoption and further development by the research community.

Summary updated

Abstract

Abstract In multisite neuroimaging studies there is often unwanted technical variation across scanners and sites. These “scanner effects” can hinder detection of biological features of interest, produce inconsistent results, and lead to spurious associations. We assess scanner effects in two brain magnetic resonance imaging (MRI) studies where subjects were measured on multiple scanners within a short time frame, so that one could assume any differences between images were due to technical rather than biological effects. We propose mica ( m ultisite i mage harmonization by C DF a lignment), a tool to harmonize images taken on different scanners by identifying and removing within-subject scanner effects. Our goals in the present study were to (1) establish a method that removes scanner effects by leveraging multiple scans collected on the same subject, and, building on this, (2) develop a technique to quantify scanner effects in large multisite trials so these can be reduced as a preprocessing step. We found that unharmonized images were highly variable across site and scanner type, and our method effectively removed this variability by warping intensity distributions. We further studied the ability to predict intensity harmonization results for a scan taken on an existing subject at a new site using cross-validation.
BibTeX
@misc{wrobel2019intensity,
  author = {Julia Wrobel and ML Martin and Rohit Bakshi and Peter A. Calabresi and Mark Elliot and D Raolf and RC Gur and RE Gur and RG Henry and Govind Nair and Jiwon Oh and Nico Papinutto and Daniel Pelletier and DS Reich and William D. Rooney and T. E. Satterthwaite and W. Stern and Karthik Prabhakaran and Nancy L. Sicotte and RT Shinohara and Jeff Goldsmith and on behalf of the NAIMS Cooperative},
  title = {{Intensity warping for multisite MRI harmonization}},
  year = {2019},
  doi = {10.1101/679357},
  howpublished = {bioRxiv},
}
Figure