Modeling T1 Resting-State MRI Variants Using Convolutional Neural Networks in Diagnosis of OCD

TL;DR

This study employs convolutional neural networks to analyze T1 resting-state MRI scans for diagnosing OCD, revealing moderate accuracy and supporting the p-factor theory of mental disorders.

Contribution

It introduces a novel application of CNNs, including ResNet50 and MobileNet, to distinguish OCD from other mental disorders using TRS-MRI data.

Findings

ResNet50 achieved 82.08% accuracy for schizophrenia.

The models had 88.75% accuracy for major depressive disorder.

OCD classification accuracy was approximately 54.4%.

Abstract

Obsessive-compulsive disorder (OCD) presents itself as a highly debilitating disorder. The disorder has common associations with the prefrontal cortex and the glutamate receptor known as Metabotropic Glutamate Receptor 5 (mGluR5). This receptor has been observed to demonstrate higher levels of signaling from positron emission tomography scans measured by its distribution volume ratios in mice. Despite this evidence, studies are unable to fully verify the involvement of mGluR5 as more empirical data is needed. Computational modeling methods were used as a means of validation for previous hypotheses involving mGluR5. The inadequacies in relation to the causal factor of OCD were answered by utilizing T1 resting-state magnetic resonance imaging (TRS-MRI) scans of patients suffering from schizophrenia, major depressive disorder, and obsessive-compulsive disorder. Because comorbid cases often occur within these disorders, cross-comparative abilities become necessary to find distinctive characteristics. Two-dimensional convolutional neural networks alongside ResNet50 and MobileNet models were constructed and evaluated for efficiency. Activation heatmaps of TRS-MRI scans were outputted, allowing for transcriptomics analysis. Though, a lack of ability to predict OCD cases prevented gene expression analysis. Across all models, there was an 88.75% validation accuracy for MDD, and 82.08% validation accuracy for SZD under the framework of ResNet50 as well as novel computation. OCD yielded an accuracy rate of around 54.4%. These results provided further evidence for the p-factor theory regarding mental disorders. Future work involves the application of alternate transfer learning networks than those used in this paper to bolster accuracy rates.