Decision processes in 3D structural MRI schizophrenia classification evaluated with saliency maps | JoVE Visualize

Area of Science:

Neuroscience
Artificial Intelligence
Psychiatry

Background:

Clinical decision support systems for psychiatric disorders like schizophrenia can be enhanced by machine learning models utilizing neuroimaging data.
Deep learning (DL) models excel at detecting complex patterns in images without predefined region information, making them suitable for objective diagnosis, prognosis, and treatment selection.
However, DL models often lack transparency in their decision-making processes, hindering clinical translation.

Purpose of the Study:

To evaluate the plausibility of seven DL architectures in schizophrenia classification using neuroimaging data via gradient-weighted class activation mapping (Grad-CAM).
To develop a method for translating Grad-CAM saliency maps into interpretable anatomical markers for schizophrenia.
To demonstrate the necessity and feasibility of explainable AI (XAI) methods in deriving biomarkers for psychiatric disorders.

Main Methods:

Qualitative and quantitative evaluation of seven DL architectures for schizophrenia classification.
Application of gradient-weighted class activation mapping (Grad-CAM) to assess model decision-making.
Development of an approach to translate saliency maps into interpretable anatomical markers.

Main Results:

Only two out of seven evaluated DL models based their schizophrenia classification decisions on plausible structural brain information, despite comparable classification performance.
The developed approach successfully translated saliency maps into interpretable anatomical markers.
Candidate regions identified corresponded to known schizophrenia markers.

Conclusions:

Explainable AI methods are essential for ensuring the clinical validity and interpretability of DL models in psychiatric neuroimaging.
The study demonstrates the feasibility of using XAI techniques to derive objective, interpretable biomarkers for schizophrenia from neuroimaging data.
This approach can aid in the clinical translation of DL applications and biomarker discovery for psychiatric disorders.