Graph-Based Encoder-Decoder for Harmonizing Structural Connectomes Across MRI Sites
Poster No:
1221
Submission Type:
Abstract Submission
Authors:
Jagruti Patel1, Mikkel Schöttner1, Thomas Bolton1, Patric Hagmann1
Institutions:
1Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL), Lausanne, Swizterland
First Author:
Jagruti Patel
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Co-Author(s):
Mikkel Schöttner
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Thomas Bolton
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Patric Hagmann
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Department of Radiology, Lausanne University Hospital and University of Lausanne (CHUV-UNIL)
Lausanne, Swizterland
Introduction:
Physical connections between pairs of brain regions form the structural connectome (SC) [2], which serves as a potential biomarker for identifying and understanding brain disorders [1]. Harmonizing SCs derived from neuroimaging data across multiple sites can enhance the generalizability and sensitivity of these findings [3]. In this work, we propose a graph neural network-based [7] encoder decoder framework to achieve this, ensuring the preservation of brain network structure while eliminating the need for traveling subjects (TS) [5].
Methods:
Our dataset included the minimally preprocessed T1-weighted imaging (T1W) and diffusion-weighted imaging (DWI) data of 1061 subjects from the Human Connectome Project Young Adult dataset [6]. 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). Then, SCs were derived by parcellating the T1W data into 274 regions and performing deterministic tractography on the DWI data to infer the number of fibers linking them [4].
The subjects were grouped into sets of around 848 for training (train), 106 for validation (val) and 107 for testing (test), accounting for family relations.
The train set consisted of independent SCs acquired with only one acquisition parameter (AP) combination. Data augmentation was performed by randomly sampling pairs of SCs and combining mutually exclusive subsets of edges from each of the two selected SCs. This yielded around 4000 matrices for training upon augmentation. Note that augmentation was also performed for similar subject pairs at other APs.
The val and test sets consisted of SCs acquired with the highest bval, highest res (bval=3000, res=1.25) (HBHR) and lowest bval, lowest res (bval=1000, res=2.3) (LBLR).
The graph encoder decoder framework was designed as shown in Figure 1. It has three parts: an encoder decoder network (EDN), a mapping network (MN) and a site classifier network. During training, weighted SC belonging to a certain site or AP was given as input as well as target to the EDN, and the condition vector corresponding to this site was given as input to the MN. During evaluation, the LBLR SC was given as input to the EDN and the condition vector corresponding to the HBHR SC as input to the MN. During both times, an identity matrix was given as the input feature matrix. The optimized loss function was the mean absolute error (MAE) between the predicted and target SCs combined with the cross entropy loss for site prediction by the classifier where the weightage of MAE compared to the cross entropy loss was slowly increasing with the number of epochs.
For validation, the input, predicted and target SCs of two test subjects were plotted as shown in Figure 2. Also, the MAE and fingerprinting accuracy were calculated between the input and target SCs and the predicted and target SCs.
The subjects were grouped into sets of around 848 for training (train), 106 for validation (val) and 107 for testing (test), accounting for family relations.
The train set consisted of independent SCs acquired with only one acquisition parameter (AP) combination. Data augmentation was performed by randomly sampling pairs of SCs and combining mutually exclusive subsets of edges from each of the two selected SCs. This yielded around 4000 matrices for training upon augmentation. Note that augmentation was also performed for similar subject pairs at other APs.
The val and test sets consisted of SCs acquired with the highest bval, highest res (bval=3000, res=1.25) (HBHR) and lowest bval, lowest res (bval=1000, res=2.3) (LBLR).
The graph encoder decoder framework was designed as shown in Figure 1. It has three parts: an encoder decoder network (EDN), a mapping network (MN) and a site classifier network. During training, weighted SC belonging to a certain site or AP was given as input as well as target to the EDN, and the condition vector corresponding to this site was given as input to the MN. During evaluation, the LBLR SC was given as input to the EDN and the condition vector corresponding to the HBHR SC as input to the MN. During both times, an identity matrix was given as the input feature matrix. The optimized loss function was the mean absolute error (MAE) between the predicted and target SCs combined with the cross entropy loss for site prediction by the classifier where the weightage of MAE compared to the cross entropy loss was slowly increasing with the number of epochs.
For validation, the input, predicted and target SCs of two test subjects were plotted as shown in Figure 2. Also, the MAE and fingerprinting accuracy were calculated between the input and target SCs and the predicted and target SCs.
Results:
As seen in Figure 2, the predicted SCs are denser and closer to the target SCs as compared to the input SCs (MAE decreased from 8.18 to 6.89). Also, fingerprinting accuracy increased from 19.6% to 98.1%.
Conclusions:
Our graph neural network-based encoder and decoder effectively harmonized SCs obtained from two completely different APs without requiring any TS. Additionally, the proposed data augmentation strategy significantly enhanced the training dataset. This model can be further validated for harmonizing brain connectomes acquired from different sites with multiple factors of variation.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
Other Methods
Novel Imaging Acquisition Methods:
Anatomical MRI
Diffusion MRI 2
Keywords:
Acquisition
Machine Learning
STRUCTURAL MRI
Tractography
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
Other - Connectome ; Graph Neural Networks ; Harmonization ; Fingerprinting ; Human Connectome Project
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
[2] Contreras, J. A., et al. (2015). The structural and functional connectome and prediction of risk for cognitive impairment in older adults. Current behavioral neuroscience reports, 2, 234-245.
[3] Onicas, A. I., et al. (2022). Multisite harmonization of structural DTI networks in children: An A-CAP study. Frontiers in Neurology, 13, 850642.
[4] Patel, J., et al. (2024). Modeling the impact of MRI acquisition bias on structural connectomes: Harmonizing structural connectomes. Network Neuroscience, 1-30.
[5] Tian, D., et al. (2022). A deep learning-based multisite neuroimage harmonization framework established with a traveling-subject dataset. NeuroImage, 257, 119297.
[6] Van Essen, D. C., et al. (2012). The Human Connectome Project: a data acquisition perspective. Neuroimage, 62(4), 2222-2231.
[7] Zhou, J., et al. (2020). Graph neural networks: A review of methods and applications. AI open, 1, 57-81.