Information has been renewed.

What is your study programme?

Presentation Master's Thesis - Enrico Erler - Psychological Research Methods

Last modified on 08-08-2022
Depression classification and brain feature identification with structural MRI using 3D convolutional neural networks and gradient-weighted class activation mapping
Show information for your study programme
You're currently viewing general information. Choose your study programme to see additional information that's specific to your study programme, such as deadlines, regulations and contact details.
What is your study programme?
Start date
11-08-2022 06:30
End date
11-08-2022 07:30

This is an online presentation. You can register here!External link

Major depressive disorder (MDD) puts a severe burden on the individual and society. For this reason, more accurate and effective diagnosis and treatment is a top priority for research and practice. One pillar of these efforts constitutes the identification of structural biomarkers. We present an explainable convolutional neural network (CNN) which may help to quickly and accurately diagnose MDD and identify potential biomarkers, giving new directions for treatment and research. 

Method: Using a sample of 1,435 subjects from two datasets, we trained a 3D CNN on structural magnetic resonance images (MRIs). The network classified MDD in both a binary (healthy control (HC) vs. MDD) and categorical (HC vs. subclinical- vs. clinical MDD) classification task. To identify potential biomarkers, we backtracked which brain regions the model deemed important using spatially aligned averaged Gradient-weighted Class Activation Maps (Grad-CAMs). Results: Our CNN achieved a 95.1% (±3.51% at 95% CI) test accuracy on binary- and 83.6% (±6.01% at 95% CI) accuracy on categorical classification. As expected, Grad-CAMs highlighted a plethora of brain regions, illustrating the potential scope and complexity of structural differences that are associated with MDD, but primarily included regions around the basal ganglia, brain stem, lateral ventricles and hippocampus. However, the Grad-CAMs demonstrated a high sensitivity to sample exclusions, possibly implying a bias through model overfitting.