Assessing the Generalizability of an HCP Connectome Harmonization Model

1507 

Abstract Submission 

Jagruti Patel¹, Yasser Alemán-Gómez¹, Sebastien Tourbier¹, Mikkel Schöttner¹, Thomas Bolton¹, Patric Hagmann¹

¹Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL), Lausanne, Switzerland

Jagruti Patel  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Yasser Alemán-Gómez  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Sebastien Tourbier  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Mikkel Schöttner  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Thomas Bolton  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Patric Hagmann  
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Switzerland

Data harmonization (DH) is crucial to mitigate heterogeneity in multi-site neuroimaging studies and thereby enhance statistical power and generalizability [9]. A factor contributing to such heterogeneity is the variability in acquisition parameters (APs) of the data collected across sites [10]. While there is no gold standard in designing a DH model [5], information regarding APs makes valuable prior knowledge to improve performance [3]. In this work, we have designed such a model to harmonize structural connectomes (SCs) [4].

Our training dataset included the minimally preprocessed T1-weighted imaging (T1W) and diffusion-weighted imaging (DWI) data of 150 subjects from the Human Connectome Project Young Adult (HCP-YA) dataset [8]. The DWI data was resampled to have two factors of variation: resolution (res, 1.25 [original] and 2.3 [downsampled] mm isotropic) and b-value (bval, 1000 and 3000 s/mm²), mimicking a training scenario with 150 "traveling subjects" [7] and 4 acquisition sites. Then, SCs were derived by parcellating the T1W data into 274 regions and performing deterministic tractography on the DWI data to infer the numbers of fibers linking them [6].

Linear regression (LR) was performed independently on each structural connection:
ŷ = β0 + β1 bval + β2 res + β3 bval*res,
where ŷ is an estimated structural connection across subjects and AP combinations, and [β0, β1, β2, β3] is the estimated weight vector for the connection at hand.

Our independent validation dataset consisted of the T1W and DWI data of 11 healthy subjects from the Lausanne Psychosis Cohort [1,2]. The DWI data included diffusion tensor imaging (DTI), bval = 0 and 1000 s/mm², res = 2 x 2 x 3 mm³ or 2 x 2 x 3.3 mm³ depending on the subject, as well as diffusion spectrum imaging (DSI), maximum (max) bval = 8000 s/mm², res = 2.2 x 2.2 x 3 mm³. The DSI data was resampled to have max bval as a factor of variation (3000, 5000 and 8000 s/mm²). SCs were derived as above except that fiber orientation distribution functions were estimated using constrained spherical deconvolution (order 6) for DTI and simple harmonic oscillator‐based reconstruction and estimation (order 6) for DSI.

The above LR model was used to harmonize DTI-derived SCs to DSI-derived SCs at the three modeled max bval settings. Mean L1 distances across connections were calculated between the DSI SCs of different subjects, and between the DTI and DSI SCs of the same subjects before and after DH (Fig. 1). DTI vs DSI cross-subject distance matrices were also calculated before and after DH (Fig. 2). Since DTI and DSI data resolutions were not isotropic, a cube root of the volume of each voxel was used as res in the above model.

For DSI at max bval = 3000 and 5000 s/mm², intra-subject distances decreased after DH and became smaller than inter-subject distances (Fig. 1). However, the opposite was observed for max bval = 8000 s/mm².

Intra-subject distances (diagonal entries in Fig. 2 matrices) decreased after DH and were the lowest for all but one subject for DSI data at max bval = 3000 and 5000 s/mm², denoting that subject-specific features required for fingerprinting were not removed by our method. However, at max bval = 8000 s/mm², differences between inter-subject and intra-subject distances were less pronounced.

Although our simple LR model was trained on the high-quality HCP-YA dataset, it generalized to a different clinical dataset. While it was trained to harmonize DTI data from bval = 1000 to 3000 s/mm², it also efficiently harmonized DTI data at bval = 1000 s/mm² to DSI data until a max bval of 5000 s/mm², demonstrating generalizability and robustness. Hence, simple models should not be disregarded completely for more complex ones, especially when it comes to clinical applications.

Connectivity (eg. functional, effective, structural) ¹

Other Methods

Anatomical MRI

Diffusion MRI ²

Acquisition

Machine Learning

STRUCTURAL MRI

Tractography

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC

Other - Connectome, Harmonization, Generalizability, Human Connectome Project, Fingerprinting