Related Experiment Video
Updated: Mar 23, 2026
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
An Efficient Contrastive Deep Learning Model for Identifying Schizophrenia-Specific Neuroanatomical Variations
Lin Du1,2, Biying Peng1,2, Yuqing Sun1,2
1State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing 100875, China.
Background And Hypothesis:
Schizophrenia (SZ) is a debilitating mental disorder characterized by heterogeneous clinical manifestations and widespread brain structural abnormalities. However, the relationship between individual neuroanatomical abnormalities and clinical symptoms remains inconsistent. We hypothesize that isolating SZ-specific neuroanatomical variations could deepen our understanding of its pathophysiology and yield more reliable biomarkers for its clinical heterogeneity.
Study Design:
To investigate this, we developed DECODE-SZ (Dual Encoder Contrastive Decoding for Schizophrenia), a novel model combining contrastive learning, 3D convolutional neural networks, and variational autoencoders (VAE) to isolate SZ-specific neuroanatomical features. We applied this model to structural MRI data from 641 patients diagnosed with SZ and 609 healthy controls across 8 independent sites in China, employing a leave-one-site-out cross-validation approach to ensure robust and unbiased results. Our analysis focused on examining the relationship between SZ-specific gray matter alterations and clinical symptoms (measured by PANSS scores), while also considering non-clinical variables such as age, sex, and education.
Study Results:
The DECODE-SZ model successfully extracted SZ-specific gray matter features, revealing that these features, rather than common variations with controls, were more strongly associated with PANSS scores. Consistent brain regions exhibiting these alterations were identified across multiple sites, supporting the reliability of the findings. A control experiment using a traditional VAE model demonstrated the superior performance of DECODE-SZ in isolating meaningful SZ-specific neuroanatomical variations.
Conclusions:
Our findings highlight the potential of SZ-specific neuroanatomical alterations as key biomarkers for clinical outcomes in SZ. DECODE-SZ offers a promising tool for advancing the understanding of SZ and may inform future diagnostic and therapeutic strategies.
