Multi-scanner Harmonization of Paired Neuroimaging Data via Structure Preserving Embedding Learning

Abstract

Combining datasets from multiple sites/scanners has been becoming increasingly more prevalent in modern neuroimaging studies. Despite numerous benefits from the growth in sample size, substantial technical variability associated with site/scanner-related effects exists which may inadvertently bias subsequent downstream analyses. Such a challenge calls for a data harmonization procedure which reduces the scanner effects and allows the scans to be combined for pooled analyses. In this work, we present MISPEL (Multi-scanner Image harmonization via Structure Preserving Embedding Learning), a multi-scanner harmonization framework. Unlike existing techniques, MISPEL does not assume a perfect coregistration across the scans, and the framework is naturally extendable to more than two scanners. Importantly, we incorporate our multi-scanner dataset where each subject is scanned on four different scanners. This unique paired dataset allows us to define and aim for an ideal harmonization (e.g., each subject with identical brain tissue volumes on all scanners). We extensively view scanner effects under varying metrics and demonstrate how MISPEL significantly improves them.

Figure 1.. Example of scanner effects.

Top: a subject’s paired axial slices, coregistered across the four different scanners (GE, Philips, Prisma, and Trio) in our multi-scanner dataset. Bottom: the corresponding intensity histograms (of whole scan) with identical axes. In this example of the paired slices, the scanner effects are immediately noticed with the varying intensity histograms.

Figure 2.. Illustration of MISPEL.

For each of j = 1 : N input scans and for each of i = 1 : M scanners, Enc_i (U-Net) outputs the corresponding latent embeddings: ${\mathbf{Z}}_i^j=En{c}_i(X_i^j)$. The corresponding Dec_i (linear function) maps the embeddings to the output: ${\overline{X}}_i^j=De{c}_i({\mathbf{Z}}_i^j)$. Step 1 Embedding Learning: Enc_i=1:M and Dec_i=1:M are updated using the embedding coupling loss (${\mathcal{L}}_{coup}$) and the reconstruction loss (${\mathcal{L}}_{recon}$). Step 2 Harmonization: Only Dec_i=1:M are updated using the harmonization loss (${\mathcal{L}}_{harm}$) and the reconstruction loss (${\mathcal{L}}_{recon}$). Refer to Alg. 1 for details on training.

Figure 3.. Volume distribution boxplots.

In our paired data (i.e., each subject scanned on multiple scanners with little biological differences), identical volume distributions are expected across the scanners.

Figure 4.. Visual assessment of a slice.

Rows and columns correspond to methods and scanners respectively.

Figure 5.. Root mean square error (RMSE) bar plots.

In our paired data, lower values of RMSE is expected which shows lower deviation of measures across scanners.

Figure 6.. Dice similarity score (DSC) bar plots.

In our paired dataset, larger values of DSC is expected as denotes more overlap of the volume segmentations between scanners.