Enhanced Functional Connectivity Representation by Contrastive Learning for Brain Disease Diagnosis

1410 

Abstract Submission 

Wonsik Jung¹, Heung-Il Suk¹

¹Korea University, Seoul, Republic of Korea

Wonsik Jung  
Korea University
Seoul, Republic of Korea

Heung-Il Suk  
Korea University
Seoul, Republic of Korea

Deep learning models utilizing resting-state functional magnetic resonance imaging (rs-fMRI) have been widely used for diagnosing psychiatric conditions like major depressive disorder (MDD) and autism spectrum disorder (ASD). However, the high costs of data collection and the inherent complex, high-dimensional nature of rs-fMRI data limits its availability, affecting both the statistical power and generalizability of these models in clinical settings [1]. To overcome these issues, existing studies focus on two main strategies: i) data augmentation and ii) transfer learning. The objective of this work is to introduce a novel framework combining multi-view region masking with a self-supervised learning approach, leveraging these strategies to enhance the diagnosis and analysis of brain disorders.

As depicted in Figure 1(a), our proposed framework is structured into two phases: i) the pretext model, which serves as the backbone network using an autoencoder (AE)-based architecture. This model is designed to learn enriched functional representation through functional connectivity (FC) from unlabeled fMRI data and ii) a fine-tuning model trained on labeled (unseen) fMRI data for brain disorder-specific tasks. To be specific, for the pretext model, we employ a random ROI-level masking strategy [2] to augment the FC matrices and then feed into the shared networks for inter-regional relation learning in a self-supervised learning manner. Henceforth, this model undergoes fine-tuning to refine functional representation for brain disorder-specific diagnosis and analysis.

We compared our proposed framework with AE-based approaches [2] and state-of-the-art methods [6,7,8]. Our method achieved superior results in 5-fold cross-validation, with an average accuracy of 73.38% (±3.01) and an AUC of 0.778 (±0.045) on the ABIDE dataset, and similarly, an accuracy of 75.95% (±2.64) and an AUC of 0.830 (±0.035) on the REST-MDD dataset, surpassing all competing methods. Furthermore, to validate the generalizability of our approach, we implemented an ablation study, which involved altering the ratio of random ROI-level masking (as shown in Figure 1(b)) and conducting cross-site classification. Lastly, we applied layer-wise relevance propagation [9] to obtain insights into the interconnections across multi-institutional datasets. As depicted in Figure 2, we observed statistically significant representative regions between groups, colored red for p<0.05 (SOG.L, MOG.L, SPG.R, ANG.R, PCUN.L, PCUN.R, PAL.R, CRBL6.R, IFGoperc.L), yellow for p<0.01 (SMA.R, OLF.R, REC.R, ACG.L, ACG.R, HIP.L, CAL.L, FFG.L, SMG.R, CRBL9.L), and green for p<0.001 (MFG.R, PAL.L).
Notably, among them, specific regions (e.g., the superior occipital gyrus (SOG), middle occipital gyrus (MOG), superior parietal gyrus (SPG), angular gyrus (ANG), precuneus (PCUN), hippocampus (HIP), anterior cingulate gyrus (ACG), middle frontal gyrus (MFG)) have been confirmed to be associated with autism.

We introduce a novel deep learning framework for diagnosing psychiatric disorders using fMRI, effectively addressing the challenges of high data collection costs and data complexity. Our framework, integrating multi-view region masking with self-supervised learning, consists of a two-phase approach: an AE-based model for training from unlabeled data and a fine-tuning model for specific brain disorders. Through extensive experiments conducted on two multi-institutional datasets, our method demonstrated the superiority of accuracy and generalizability compared to comparative methods, with layer-wise relevance propagation providing further insights into critical brain regions for diagnosis.

Neurodevelopmental/ Early Life (eg. ADHD, autism)

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ¹

fMRI Connectivity and Network Modeling ²

Autism

Data analysis

DISORDERS

FUNCTIONAL MRI

Machine Learning

Psychiatric Disorders