Interpretable Deep Learning for Multiclass Psychological Disorder Classification Using CNN and TCAV Using fMRI

Title

Interpretable Deep Learning for Multiclass Psychological Disorder Classification Using CNN and TCAV Using fMRI

Creator

Nagashruthi, M.K.; Hemanth, K.S.

Description

In this paper a 3D convolutional neural network model is presented that has been improved with squeeze-and-excitation blocks and residual blocks to categorize functional MRI data across healthy controls, schizophrenia, and attention deficit hyperactivity disorder classes. A post hoc explainability framework was applied to concept activation vectors based on mean activation for each class and testing using TCAV. The merging of CAV and TCAV for fMRI data was to improve transparency by providing an interpretable model. This helps in understanding the prediction sensitivity of the models. Concept vectors are defined through the extraction and analysis of intermediate activations from the CNN-SE model. These vectors are then used to calculate the TCAV scores, which indicate the degree to which each brain region influences the model output. The precision of the deep learning model was up to 79%. Similarity matrices indicate the degree of overlap and correlate with the model's results. Visualizations such as activation heat maps and glass brain overlays based on the AAL atlas further support the interpretability of the model, making it more transparent and suitable for clinical applications. Variations in activation between classes can be observed using a visual feature plot. This framework allows mapping model predictions to interpretable neuroanatomical regions and identifies classspecific dependencies on particular brain networks. 2025 IEEE.

Source

Conference Proceedings - 2025 IEEE 4th International Conference on Data, Decision and Systems, ICDDS 2025;pp.147-152

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

classification; concept vector; explainability; fMRI; interpretability

Coverage

Nagashruthi M.K., School of Computer Science and Applications, Reva University, Bengaluru, India; Hemanth K.S., Christ University, Department of Computer Science, Bengaluru, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833155479-8;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/ICDDS67737.2025.11344721

https://www.scopus.com/pages/publications/105033365376?origin=resultslist