Dual-representation structural MRI classification of psychiatric disorders using deep learning and large language...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Psychiatry

Background:

• Accurate differentiation between schizophrenia and bipolar disorder is challenging due to overlapping symptoms and subtle neuroanatomical differences.
• Current diagnostic methods can be subjective and lack objective biomarkers.
• Structural MRI offers potential for objective assessment but requires advanced analytical techniques.

Purpose of the Study:

• To develop and evaluate a dual-representation structural MRI framework for improved classification of psychiatric disorders.
• To compare the diagnostic utility of raw MRI slices versus tissue segmentation maps using deep learning.
• To leverage Large Language Models (LLMs) for enhanced interpretability of neuroimaging findings.

Main Methods:

• Utilized a dual-representation framework analyzing raw T1-weighted MRI slices and color-coded tissue segmentation maps.
• Employed two independently trained ResNet-18 Convolutional Neural Networks (CNNs) for feature extraction.
• Applied Transfer Learning (TL) and Domain Adaptation (DA) techniques to a dataset of 103 subjects across four groups (Healthy Controls, Schizophrenia Spectrum, Bipolar Disorder with Psychosis, Bipolar Disorder without Psychosis).
• Integrated a Large Language Model (LLM) for post-hoc analysis and interpretation of CNN outputs.

Main Results:

• The dual-representation approach demonstrated improved four-way classification performance compared to single-representation methods.
• Systematic comparison revealed differential contributions of raw versus segmentation-based MRI inputs to classification accuracy.
• LLM-assisted interpretation provided insights into the neuroanatomical features driving diagnostic predictions.

Conclusions:

• The proposed dual-representation framework enhances diagnostic classification accuracy in psychiatric disorders.
• Combining deep learning with LLM interpretability offers a promising avenue for transparent and informative psychiatric neuroimaging tools.
• This approach has the potential to support more objective and reliable psychiatric diagnoses.