Plasma CycleGAN: Integrating Blood‐based Biomarkers for Cross‐modality Translation from MRI to PET

Yanxi Chen, Yi Su, Yu Fu, Kewei Chen, David Weidman, Richard J. Caselli, Eric M. Reiman, Yalin WangAlzheimer 2024 · DOI

TL;DR

This paper introduces Plasma CycleGAN, a novel generative model that synthesizes PET images from MRI scans by incorporating blood-based biomarkers (BBBMs) and clinical data. This approach significantly improves the accuracy of MRI-to-PET translation compared to existing methods, paving the way for more comprehensive and cost-effective neuroimaging analysis.

What Problem Does This Paper Address?

Medical imaging often relies on multiple modalities, like MRI and PET, to provide complementary information. However, acquiring both types of scans can be expensive and time-consuming. MRI offers excellent structural detail, while PET excels at visualizing metabolic activity or the presence of specific molecules (like amyloid plaques in Alzheimer's disease). The challenge lies in translating images from one modality to another, a process known as cross-modality translation. Existing methods, often based on Generative Adversarial Networks (GANs) like CycleGAN, have shown promise in generating PET images from MRI. However, these methods haven't fully explored how readily available blood-based biomarkers (BBBMs) – which are known to correlate with PET findings – can enhance this translation. This paper aims to bridge this gap by developing a method that leverages BBBMs and other clinical data to improve the accuracy and reliability of synthesizing PET images from MRI.

Methods

The core of this work is Plasma CycleGAN, a generative model built upon the CycleGAN architecture. The key innovation is the integration of BBBMs and other clinical covariates (like age and sex) as additional conditions within the input domain of the GAN. This is achieved by introducing a "masked perturbation" into the MRI image, where the voxel-level magnitudes are learned by the neural network. This allows the model to condition its PET image generation not only on the structural MRI but also on the patient's biological and clinical profile. The authors also mention improvements to the backbone of the CycleGAN architecture to better handle these integrated conditions.

Datasets

The study utilizes data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. This dataset is particularly valuable as it includes a cohort of 160 subjects with matching:

  • Blood-based biomarkers (BBBMs)
  • T1-weighted MRI images
  • Amyloid PET scans
  • Clinical features (age, sex, weight, etc.)

This comprehensive dataset allows for the evaluation of the model's performance under various input conditions, including MRI alone, MRI combined with clinical features, and MRI combined with BBBMs.

Key Insights and Ideas

  • BBBMs are powerful predictors: The study confirms previous findings that BBBMs strongly correlate with PET-measured brain amyloid status.
  • Integrating BBBMs enhances translation: The most significant insight is that incorporating BBBMs into the MRI-to-PET synthesis process leads to demonstrably better results than using MRI alone or MRI with other clinical features. This suggests that BBBMs provide crucial biological context that helps the GAN generate more accurate PET images.
  • Plasma CycleGAN is a first: This work is highlighted as the first generative model to integrate BBBMs into a conditional cross-modality translation from MRI to PET.

Key Results

The Plasma CycleGAN model, when conditioned on MRI and amyloid biomarkers (Aβ42/40), achieved the best performance across several key metrics:

  • Structural Similarity Index (SSIM): 0.72 ± 0.16
  • Peak Signal-to-Noise Ratio (PSNR): 22.8 ± 2.3

These values represent the highest scores among all tested input combinations, indicating superior image synthesis quality. While MRI+Sex achieved the best result for Mean Squared Error (MSE), the overall performance, particularly in capturing structural similarity and signal fidelity, was superior with the inclusion of BBBMs. Pearson Correlation Coefficients (PCCs) for masked SUVR ratios ranged from 0.44 to 0.5, suggesting a good correlation between synthesized and actual PET measures.

Limitations and Future Work

The authors acknowledge that the study evaluated the model using a specific set of BBBMs and clinical covariates. Future work could explore the impact of a wider range of BBBMs and clinical features, as well as investigate the model's generalizability to different PET tracers or disease populations. Further research could also focus on optimizing the "masked perturbation" mechanism and exploring alternative GAN architectures for even greater translation accuracy. The current PCC values, while good, suggest there is still room for improvement in precisely replicating quantitative PET measures.

Summary updated

Abstract

Abstract Background Cross‐modality translation between MRI and PET presents significant challenges due to the distinct mechanisms underlying those modalities. Currently, the leading method for this translation is CycleGAN. Previous studies have proved a strong correlation between BBBMs and brain amyloid status measured by PET. However, the impact of blood‐based biomarkers (BBBMs) on PET image synthesis has not yet been thoroughly evaluated. In this study, we propose Plasma CycleGAN, a generative model based on CycleGAN to synthesize PET images from MRI, while also incorporating BBBMs and other clinical covariates. We show that Plasma CycleGAN outperforms the state‐of‐the‐art CycleGAN for MRI‐to‐PET translation. Our method is the first to integrate BBBMs in a conditional cross modality translation from MRI to PET. Method Our model was built on CycleGAN architecture, while the backbone was improved by integrating BBBMs as additional conditions in the input domain. To be specific, we introduce a masked perturbation in the MRI image, whose voxel‐level magnitudes are learnable parameters in the neural network. The dataset we used to evaluate our model is downloaded from the ADNI database. The cohort contains 160 subjects with matching BBBM, T1‐weighted MRI images, amyloid PET, and clinical features, including age, sex, weights, etc. The synthesized PET images are compared with the ground truth PET images for performance evaluation using traditional metrics like structural similarity index (SSIM), peak‐signal‐to‐noise ratio (PSNR) and mean square error (MSE). In addition, the Pearson correlation coefficients (PCCs) of the masked SUVR ratios were calculated. Result We evaluated the baseline CycleGAN model using MRI input, as well as combined inputs of MRI+BBBM and other clinical features. Our model achieved an SSIM of 0.72±0.16 and PSNR of 22.8±2.3 in PET image synthesis task using MRI+ Aβ42/40, both yielding the best results among all input combinations. MRI+sex achieved the best result using MSE measurement. PCCs of all experimental results ranged from 0.44 to 0.5. Conclusion We introduce Plasma CycleGAN, the first generative model to integrate BBBMs and clinical features in PET image synthesis from MRI images. In PET image synthesis task, our proposed integrated model outperforms the baseline CycleGAN model.
BibTeX
@inproceedings{chen2024plasma,
  author = {Yanxi Chen and Yi Su and Yu Fu and Kewei Chen and David Weidman and Richard J. Caselli and Eric M. Reiman and Yalin Wang},
  title = {{Plasma CycleGAN: Integrating Blood‐based Biomarkers for Cross‐modality Translation from MRI to PET}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.095616},
}
Figure