Individualized Brain MRI Synthesis from a Single Scan: Applications to Aging and Alzheimer's Disease

1859 

Abstract Submission 

Jingru Fu¹, Yuqi Zheng², Daniel Ferreira³, Rodrigo Moreno⁴

¹KTH, Huddinge, STOCKHOLM, ²KTH, Stockholm, stockholm, ³Karolinska Institutet, Stockholm, Sweden, ⁴KTH, Huddinge, stockholm

Jingru Fu  
KTH
Huddinge, STOCKHOLM

Yuqi Zheng  
KTH
Stockholm, stockholm

Daniel Ferreira  
Karolinska Institutet
Stockholm, Sweden

Rodrigo Moreno  
KTH
Huddinge, stockholm

Simulation of different future appearances of the MRI scans of an individual can provide relevant information to the physicians to select the most appropriate health care. In the last few years, different deep learning- (DL) based medical image generation (MIG) models have been proposed for this task. For example, generative adversarial networks (GANs) have been used for modeling the aging process and the progression of AD [1,2]. Unfortunately, these methods make important simplifications to reduce their computational costs, such as simulating a single slice per subject or downsampling original images. More importantly, GAN-based methods usually lack anatomical plausibility in generated images due to the absence of biologically informed constraints. As an alternative, we used diffeomorphic registration in [3] for this task. The main advantage of our approach compared to GANs is that the anatomical plausibility of the synthetic images is guaranteed by design. However, the main drawback of our method is that it requires two images per subject.

In this study, we introduce Individualized Brain Synthesis (InBrainSyn), a MIG to synthesize high-resolution longitudinal MRI scans that faithfully replicate subject-specific neurodegeneration in AD and aging based on a single scan.

In this study, we used T1-weighted (T1w) MRI from OASIS-3 dataset [4]. Fig.1, shows the proposed pipeline. First, T1w MRI scans were preprocessed and partitioned into cognitive normal (CN) and AD cohorts. Second, AtlasGAN [5], a DL-based template generation method, was employed to produce age-resolved high-quality templates for both the CN and AD cohorts. AtlasGAN uses a diffeomorphic registration model that estimates a stationary velocity field (SVF) that is used to create a diffeomorphic deformation field through an integration layer and a spatial transformer network (STN) [6]. Third, registration was used to estimate SVFs between every follow-up template and age-matched template of the given subject image, independently for the two cohorts. Also, an SVF was estimated between the image of the subject and its age-matched template. Then, parallel transport [7] was used to translate the SVFs from the templates to the SVF computed for the specific subject. This way, the trajectory of the cohort is translated to create the subject-specific trajectory. Lastly, STN and integration were applied to the subject-specific SVFs to generate images that simulate morphological atrophy over time due to normal aging, AD, or both, depending on the used templates.

We used 1,893 T1w MRI scans from 1,020 participants to train the cohort-level models and a separate test set of 473 MRI scans from 398 participants to evaluate it. Fig. 2 shows a comparison between the acquired images and the synthetic images generated using the CN template (InBrainSyn_CN), AD template (InBrainSyn_AD), and both (InBrainSyn_Inter). In the latter case, we use the CN template when the subject had a CDR=0 and the AD template for CDR>0. For comparison, we also show images generated with Simul@trophy [8]. As shown, InBrainSyn_Inter generates sharper images with a closer appearance to the ground truth. We also quantified the quality of synthetic images using two conventional similarity criteria, the structural similarity index (SSIM) and normalized cross-correlation (NCC). The results show the proposed InBrainSyn is superior to a baseline method in both the CN cohort (SSIM: 0.91 (±0.06) vs. 0.80 (±0.02); NCC: 0.98 (±0.02) vs. 0.94 (±0.01)) and the AD cohort (SSIM: 0.91 (±0.04) vs. 0.80 (±0.02); NCC: 0.98 (±0.01) vs. 0.93 (±0.01)).

The proposed framework can synthesize individualized images that simulate normal aging or AD courses from a single scan. The main advantages of the framework are that it is efficient, the generated images are anatomically plausible by design, and it is flexible to simulate transitions from a healthy state to AD, as shown in Fig.2.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Aging

Image Registration and Computational Anatomy

Methods Development ¹

Anatomical MRI

Aging

Degenerative Disease

Machine Learning

STRUCTURAL MRI