Temporal-Regional Pre-training Model for ADHD Diagnosis via rs-fMRI Compatible with Any Brain Atlas
Poster No:
1130
Submission Type:
Abstract Submission
Authors:
Da Woon Heo1, Kwanseok Oh1, Hojin Jang2, Heung-Il Suk1
Institutions:
1Korea University, Seoul, Republic of Korea, 2Korea University and Massachusetts Institute of Technology, Seoul, Republic of Korea
First Author:
Co-Author(s):
Introduction:
Advances in brain imaging and deep learning have greatly improved our understanding of brain functions, enabling impactful applications such as the diagnosis of neurological diseases and integration with robotics. For example, BCI leverages brain imaging to support neurological rehabilitation, enhancing communication, decision-making, and motor coordination. Among the various neuroimaging modalities, rs-fMRI is a leading tool for studying brain function and diagnosing diseases.
ADHD is a disorder affecting children and adolescents, impacting social interactions, family relationships, and overall well-being (Polanczyk et al., 2007). Diverse rs-fMRI data analysis methods are available for diagnosing or analyzing brain diseases. These include voxel-level BOLD signals, atlas-based parcellated signals, and functional connectivity metrics. While voxel-level analysis provides high spatial resolution, it is computationally intensive. Atlas-based parcellation simplifies processing but risks losing temporal dynamics, and FC approaches often miss intrinsic temporal characteristics of the BOLD signal.
Current models heavily depend on specific brain atlases, requiring retraining for new atlas data. To address these limitations, we propose a Transformer-based model integrating region-wise temporal representation and a network-wise module. This design adapts to various atlases without retraining, offering robust, efficient, and generalizable performance across diverse configurations.
ADHD is a disorder affecting children and adolescents, impacting social interactions, family relationships, and overall well-being (Polanczyk et al., 2007). Diverse rs-fMRI data analysis methods are available for diagnosing or analyzing brain diseases. These include voxel-level BOLD signals, atlas-based parcellated signals, and functional connectivity metrics. While voxel-level analysis provides high spatial resolution, it is computationally intensive. Atlas-based parcellation simplifies processing but risks losing temporal dynamics, and FC approaches often miss intrinsic temporal characteristics of the BOLD signal.
Current models heavily depend on specific brain atlases, requiring retraining for new atlas data. To address these limitations, we propose a Transformer-based model integrating region-wise temporal representation and a network-wise module. This design adapts to various atlases without retraining, offering robust, efficient, and generalizable performance across diverse configurations.
Methods:
The proposed framework comprises three stages (Figure 1): two pre-training steps-Temporal Representation Learning and Network-Region Relational Learning-followed by a downstream task. The first step employs weak (i.e., scaling) and strong (i.e., jittering and permutation) augmentation (Eldele et al., 2023) for orthogonal-based learning to extract unique temporal representations for each ROI. The second step focuses on capturing functional relationships within and between networks across different atlases. This is achieved using similarity-based contrastive learning (e.g., the same main network from different atlases as positive samples and different networks from the same or different atlases as negative samples) and classification-based strategies, where the main network is masked, and the task is to classify which network is masked. Finally, the learned features are applied to diverse downstream tasks via a shared projection layer architecture.
·Figure 1. The framework of the proposed method
Results:
We used the ADHD-200 Global Competition dataset (934 subjects, including 355 ADHD subjects), and we preprocessed using C-PAC with default settings. The Yeo17 atlas and the Schaefer 2011 atlas with parcellations of 100, 200, and 300 regions were employed for pre-training. The dataset was divided into a 5-fold cross-validation.
We evaluated our proposed model for downstream tasks using five commonly used brain atlases: CC200, Schaefer100, Schaefer200, Yeo17, and AAL3. To facilitate a straightforward comparison with the state-of-the-art model BrainNT (Kan et al., 2022), we focused on the AAL3, CC200, Schaefer100, and Schaefer200 atlases owing to the inevitable dimensional constraints associated with given BrainNT's network hyperparameter properties.
Across all atlases, we observed our proposed model consistently outperformed BrainNT, as reported in Table 1. Most remarkably, on the CC200 atlas, our model achieved an AUC of 0.71 compared to BrainNT's 0.66 and an ACC of 65.09% compared to BrainNT's 63.59%. We further observed similar trends to those of other atlases.
We evaluated our proposed model for downstream tasks using five commonly used brain atlases: CC200, Schaefer100, Schaefer200, Yeo17, and AAL3. To facilitate a straightforward comparison with the state-of-the-art model BrainNT (Kan et al., 2022), we focused on the AAL3, CC200, Schaefer100, and Schaefer200 atlases owing to the inevitable dimensional constraints associated with given BrainNT's network hyperparameter properties.
Across all atlases, we observed our proposed model consistently outperformed BrainNT, as reported in Table 1. Most remarkably, on the CC200 atlas, our model achieved an AUC of 0.71 compared to BrainNT's 0.66 and an ACC of 65.09% compared to BrainNT's 63.59%. We further observed similar trends to those of other atlases.
·Table 1. Performance comparison of BrainTF (Kan et al., 2022) and the proposed model across different atlases
Conclusions:
Our proposed pre-training network significantly advances ADHD diagnosis by addressing key challenges in feature extraction and model adaptability. Its superior performance and flexibility across atlases set the stage for future developments in neuroscience research. Expanding into a foundation model with enhanced neuroscientific insights will pave the way for broader applications, making it a valuable tool for understanding and diagnosing brain disorders.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism)
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Novel Imaging Acquisition Methods:
BOLD fMRI 2
Keywords:
Attention Deficit Disorder
Data analysis
FUNCTIONAL MRI
Other - Deep Learning
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:
Provide references using APA citation style.
2. Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C. K., Li, X., & Guan, C. (2023). Self-supervised contrastive representation learning for semi-supervised time-series classification. IEEE Transactions on Pattern Analysis and Machine Intelligence.
3. Kan, X., Dai, W., Cui, H., Zhang, Z., Guo, Y., & Yang, C. (2022). Brain network transformer. Advances in Neural Information Processing Systems, 35, 25586-25599.
4. ADHD-200 Consortium. (n.d.). ADHD-200 Global Competition dataset. Retrieved from https://fcon_1000.projects.nitrc.org/indi/adhd200/
Acknowledgments: This research was supported by the MOTIE (Ministry of Trade, Industry, and Energy) in Korea, under the Fostering Global Talents for Innovative Growth Program related to Robotics (P0017311) supervised by the Korea Institute for Advancement of Technology (KIAT) and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University)).