DISARM++: Beyond scanner-free harmonization

TL;DR

DISARM++ directly harmonizes 3D T1-weighted MRIs by disentangling anatomical structure from scanner effects into latent spaces $\mathcal{B}$ and $\mathcal{S}$, enabling two harmonization modes: scanner-free and reference-scanner space. The model generalizes to unseen scanners, removes skull-stripping requirements, and maintains full-head information, enhancing downstream analyses. Across brain-age prediction ($R^{2}\approx 0.60$), AD diagnosis ($AUC\approx 0.95$), and MCI/AD diagnosis tasks, DISARM++ outperformed STGAN and IGUANe, with robust ablations confirming all components are beneficial. It offers a practical, scalable solution for cross-site neuroimaging, avoiding retraining on new datasets while improving biomarker reliability and analysis accuracy.

Abstract

Harmonization of T1-weighted MR images across different scanners is crucial for ensuring consistency in neuroimaging studies. This study introduces a novel approach to direct image harmonization, moving beyond feature standardization to ensure that extracted features remain inherently reliable for downstream analysis. Our method enables image transfer in two ways: (1) mapping images to a scanner-free space for uniform appearance across all scanners, and (2) transforming images into the domain of a specific scanner used in model training, embedding its unique characteristics. Our approach presents strong generalization capability, even for unseen scanners not included in the training phase. We validated our method using MR images from diverse cohorts, including healthy controls, traveling subjects, and individuals with Alzheimer's disease (AD). The model's effectiveness is tested in multiple applications, such as brain age prediction (R2 = 0.60 \pm 0.05), biomarker extraction, AD classification (Test Accuracy = 0.86 \pm 0.03), and diagnosis prediction (AUC = 0.95). In all cases, our harmonization technique outperforms state-of-the-art methods, showing improvements in both reliability and predictive accuracy. Moreover, our approach eliminates the need for extensive preprocessing steps, such as skull-stripping, which can introduce errors by misclassifying brain and non-brain structures. This makes our method particularly suitable for applications that require full-head analysis, including research on head trauma and cranial deformities. Additionally, our harmonization model does not require retraining for new datasets, allowing smooth integration into various neuroimaging workflows. By ensuring scanner-invariant image quality, our approach provides a robust and efficient solution for improving neuroimaging studies across diverse settings. The code is available at this link.

Figures (16)

• Figure 1: From left to right, we have a) the brain encoder $E_b$, b) the scanner encoder $E_s$, c) the generator $G$, d) the brain discriminator $D_b$ and e) the scanner discriminator $D_s$.
• Figure 2: A high-level functional diagram of the model training procedure. The encoders and the generator are single modules, despite being depicted in different colors. The varying colors for the encoders highlight the specific acquisition scanner domain of the images they process. For the generator, the different colors emphasize the scanner domain to which it is transferring the image. The images used to illustrate the procedure are acquired with the Prisma (purple) and Gyroscan Intera (green) scanners.
• Figure 3: Training procedure part related to the scanner-free loss. The images used to illustrate the procedure are acquired with the Prisma (purple) and Gyroscan Intera (green) scanners.
• Figure 4: Inference to a reference scanner for a new image $\bm{x}_{\text{j}} \in \mathcal{X}_{\text{j}}$. Specifically, the figure illustrates the transfer of an image acquired with the Skyra Fit (pink) scanner to the reference Gyroscan Intera (green).
• Figure 5: Inference to scanner-free for a new image $\bm{x}_{\text{j}} \in \mathcal{X}_{\text{j}}$. Specifically, the figure illustrates the transfer of an image acquired with the Skyra Fit (pink) scanner to the scanner-free space (black).