Huntington’s disease (HD) is a neurodegenerative disorder that progressively affects motor and cognitive functions. It is known that the disease is caused by CAG repeat expansion in the gene Hungtingtin (HTT), and is directly related to the CAG-age product (CAP score)^1,2. Previous MRI studies have shown structural degeneration in basal ganglia and a range of grey and white matter structures^3,4. However, the spatial and temporal order of the structural changes is not fully understood. In this study, we adapted a novel statistic model, namely the change point analysis⁵, to capture the critical turning points of brain volumes against disease progression, based on whole brain parcellation of T1-weighted images.

The data used in this study are from the multicenter PREDICT-HD study^4,6,7. MPRAGE data from 412 subjects were used, including 121 controls and 291 premanifest HD with CAP score ranged 140-586. The image were acquired around 1×1×1.5 mm³ resolution with heterogeneous protocols⁴. The images were segmented using our online resource MriCloud^8,9 (www.mricloud.org), which provides a fully automated T1 segmentation pipeline based on diffeomorphic image registration and multi-atlas fusion^10,11.

A two-step multivariant linear statistical model was established to examine volumetric changes in relation to the CAP score, which is a reliable indicator of the disease severity and reflects the temporal order. In the first step, the group difference between the HD subjects and controls is estimated based on the group label (g, 1 for HD and 0 for control), CAP score (c), age (a), sex (s), and intracranial volume (i).

Model 1: $$$y_{j}=\alpha+\beta\cdot c_{j}\cdot g_{j}+\gamma\cdot g_{j}+\delta\cdot a_{j}+\eta\cdot s_{j}+\lambda\cdot i_{j}+\epsilon_{j}$$$

where y denotes the volume, j is the subject index, α is a constant, and ε is the noise. The model parameters [α, β, γ, δ, η, λ] are estimated using maximum likelihood estimation. A group difference is detected when β ≠ 0 and γ ≠ 0. In the second step, only the HD population is involved, and the change point is modeled as an additional linear change that takes effect when CAP score is greater than the change point Δ, in addition to the baseline change.

Model 2: $$$y_{j}=\alpha+\beta\cdot c_{j}+\beta^{'}\cdot c_{j}\cdot H(c_{j}-\triangle)+\eta\cdot s_{j}+\lambda\cdot i_{j}+\epsilon_{j}$$$

where H is a Heaviside function such that H(t) = 1 if t>0 and H(t) = 0 otherwise. The optimal change point is obtained at Δ* when the likelihood L(Δ*) is maximal among all Δs between 200-500. A significant change point is found when β’ ≠ 0 at Δ*. The significance is evaluated by permutation test¹², and the variance of the change point is estimated via bootstrap⁵.

Figure 1 showed the change points detected in several key structures that had significant volumetric reduction as the CAP score increased. The volumes were represented as the percentage differences with respect to the controls. Note that, if a significant group difference is detected from model 1 but no significant change point is detected from model 2, we assume a linear change with CAP and assign the earliest change point (Δ*=150, in unit of CAP) for that structure, such as the basal ganglia. We mapped the change points with familywise significance of p<0.05 onto a T1-weighted brain template in Figure 2. The map revealed the earliest volumetric changes in the basal ganglia and ventricle, followed by the white matter and then the occipital and frontal cortices, while the posterior white matter changed ahead of the anterior white matter. Figure 3 showed the rate of percentage volume change after the change point, which suggested a rapid volume loss in the basal ganglia and a drastic expansion in ventricle. In addition, the anterior brain structures showed a relatively faster atrophy compared to the posterior brain.While much is known about HD and its MRI markers at discrete time points and individual structures, in this study, we mapped the spatiotemporal changing pattern of whole brain atrophy using a novel statistical model. Our findings agree with the general understanding of injury progression from the deep gray matter to the white matter and then to the cortical regions. At the same time, the data suggest a posterior-to-anterior gradient, where the posterior brain structures change earlier but at a relatively slow rate and vice versa in the anterior brain. Although the data we used in this study are not longitudinal, this large population of cross-sectional data analysis provides important knowledge about the temporal order of disease progression. In addition to the volumetric analysis, measurements from diffusion MRI or other imaging contrasts could provide more sensitive markers, and a multi-modality study is undergoing.