DISARM++: Beyond scanner-free harmonization (2505.03715v1)

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.

DISARM++: Scanner-Free Harmonization of MRI Images

Harmonization of Magnetic Resonance Imaging (MRI) data across diverse scanners is a challenging yet essential task in neuroimaging studies. Variability between different scanners can introduce significant inconsistencies, affecting the repeatability and reproducibility of extracted biomarkers and subsequent analyses. The paper "DISARM++: Beyond scanner-free harmonization" introduces a novel method for harmonizing T1-weighted MR images, targeting inter-scanner variability directly at the image level rather than post-image feature extraction. This approach is conceptualized to extend the capabilities of the previously established DISARM framework.

Methodology

DISARM++ employs an image-to-image translation model using a cycle GAN architecture, a sophisticated framework allowing adaptation to multiple scanner characteristics and configurations. Key features of the DISARM++ model include:

• Anatomical Structure and Scanner Disentanglement: The model extracts anatomical features separately from scanner-specific information using dual encoders.
• Scanner-Free Harmonization: A unique scanner-free space is created by removing scanner-related effects, introducing Gaussian noise to replace scanner-specific information. This enables the adjustment of images to a neutral, standardized appearance across different scanners.
• Cycle Consistency: Through cyclic reconstructions within the GAN framework, the model ensures that anatomical structures remain consistent between original and harmonized images.
• Attention Mechanism: Enhanced attention layers improve the capability of the model to maintain high-frequency data relevant to anatomical structures during transformation.

Strong Numerical Results

The authors tested their model using various datasets, including those containing healthy individuals, patients with Alzheimer's disease (AD), and traveling subjects. Their findings indicate substantial improvements over baseline and other state-of-the-art harmonization methods in several metrics:

• Brain Age Prediction: Achieving a coefficient of determination $R^2 \approx 0.60 \pm 0.05$, illustrating reliable predictions compared to other models.
• AD Classification: Test accuracy approximating 0.86 with a variance of ±0.03, indicating improved classification reliability.
• Diagnosis Prediction: An impressive Area Under Curve (AUC) value of $\approx 0.95$, demonstrating superior detection capabilities for distinguishing between AD and MCI due to AD.

Implications and Future Directions

DISARM++ sets a significant precedent in the field by effectively eliminating the need for extensive preprocessing steps like skull-stripping, thus preserving complete head information in MR images. Moreover, it offers robust generalization capabilities for unseen scanners, broadening potential applications and integration into diverse neuroimaging workflows. This augmentation in harmonization processes opens doors to improved biomarker extraction efficacy and presents opportunities to explore comprehensive datasets without losing critical anatomical information.

Future advances could explore the application of the DISARM++ framework to other diseases and imaging modalities. Exploring data from varied pathologies may further ensure adaptability and strengthen insights offered by harmonized datasets. Continuous improvements in harmonization techniques hold promise for enhancing the reliability of large-scale neuroimaging studies, profoundly impacting clinical and research practices.

DISARM++ is a pivotal contribution to the quest for seamless integration and standardization within neuroimaging, warranting attention to this innovative approach and its potential for broader applications and impacts.