Multi modality fusion transformer with spatio-temporal feature aggregation module for psychiatric disorder diagnosis

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Bipolar disorder (BD) diagnosis relies on identifying recurrent depressive and mild manic episodes.
• Current diagnostic methods often lack sufficient accuracy, necessitating advanced analytical approaches.
• Resting-state functional magnetic resonance imaging (rs-fMRI) offers a promising avenue for objective BD assessment.

Purpose of the Study:

• To develop a novel framework, Spatio-temporal Feature Fusion Transformer (STF2Former), for enhanced bipolar disorder diagnosis.
• To improve upon existing methods by better capturing the complex spatio-temporal dynamics in rs-fMRI data.
• To achieve higher accuracy in clinical diagnosis of bipolar disorder.

Main Methods:

• Proposed the Spatio-temporal Feature Fusion Transformer (STF2Former) framework, building upon MFFormer.
• Introduced a Spatio-temporal Feature Aggregation Module (STFAM) to extract temporal and spatial features from rs-fMRI data.
• Decoupled temporal and spatial dimensions for separate feature extraction and employed intra-modality attention for information fusion.

Main Results:

• The Spatio-temporal Feature Aggregation Module (STFAM) demonstrated effectiveness in extracting crucial features from rs-fMRI.
• STF2Former significantly outperformed the previous MFFormer model in diagnostic accuracy.
• The proposed method achieved superior results compared to other state-of-the-art approaches for bipolar disorder detection.

Conclusions:

• STF2Former offers a significant advancement in the accurate diagnosis of bipolar disorder using rs-fMRI.
• The integration of spatio-temporal feature learning enhances the model's ability to differentiate BD from other conditions.
• This framework holds potential for improving clinical decision-making in psychiatry.