3D CNN Classification Model to Identify Bipolar Disorders and Major Depressive Disorders by Rs-fMRI

503 

Abstract Submission 

Jaimie Yeh¹, Albert Yang^1,2,3,4

¹Institute of Brain Science, National Yang Ming Chiao Tung University, Taipei City, Taiwan, ²Digital Medicine and Smart Healthcare Research Center, National Yang Ming Chiao Tung University, Taipei City, Taiwan, ³Department of Medical Research, Taipei Veterans General Hospital, Taipei City, Taiwan, ⁴Brain Research Center, National Yang Ming Chiao Tung University, Taipei City, Taiwan

Jaimie Yeh  
Institute of Brain Science, National Yang Ming Chiao Tung University
Taipei City, Taiwan

Albert Yang  
Institute of Brain Science, National Yang Ming Chiao Tung University|Digital Medicine and Smart Healthcare Research Center, National Yang Ming Chiao Tung University|Department of Medical Research, Taipei Veterans General Hospital|Brain Research Center, National Yang Ming Chiao Tung University
Taipei City, Taiwan|Taipei City, Taiwan|Taipei City, Taiwan|Taipei City, Taiwan

Bipolar disorder (BD) and major depressive disorder (MDD) are prevalent psychiatric disorders that present similar clinical symptoms during depressive episodes. Distinguishing between depressed BD and MDD based on clinical symptoms is a considerable challenge, which makes the clinical treatment more difficult. Consequently, biomarkers that are capable of accurately discerning between these two disorders assume paramount importance in clinical diagnostics. Resting-state functional magnetic resonance imaging (rs-fMRI) provides diagnostic markers for studying affective disorders. Although several rs-fMRI studies have reported abnormal functional connectivity between individuals with BD and those with MDD, most have focused on identifying relevant brain alterations at the group level. Machine learning techniques offer significant practical value in predicting diagnoses and clinical outcomes at the individual level. Furthermore, deep learning applications, such as the three-dimensional convolutional neural network (3D CNN) classification model, provide a more comprehensive approach by incorporating spatio-temporal features. In this study, we aimed to utilize 3D CNN methodologies to identify neuroimaging feature differences between individuals with BD and individuals with MDD, which can help in clinical diagnosis and treatment strategies.

Rs-fMRI data were obtained from the Taiwan Aging and Mental Illness cohort, encompassing 99 individuals with BD (age: 50.66 ± 10.35; 34% in males), 90 individuals with MDD (age: 54.11 ± 13.57; 33% in males), and 193 healthy controls (HCs; age: 50.07 ± 11.37; 35% in males). The whole brain voxelwise functional connectivity map was computed for each participant. We applied z-transformation to improve the normality of the data. Subsequently, 90 functional connectivity maps for each participant were delineated based on the Automated Anatomical Labeling (AAL) atlas. The 3D characteristics of each functional connectivity map were retained for subsequent model training. Hierarchical 3D CNN classification models were applied to each functional connectivity map. Initially, we classified between the HCs and individuals with affective disorders (i.e., BD and MDD). Next, we performed 3D CNN classification models to differentiate between individuals with BD and those with MDD. Based on the outcomes of these two steps, we were able to identify critical brain features for the differential diagnosis of affective disorders.

In the classification of HCs and individuals with affective disorders, our results revealed that 3D CNN models, constructed using brain regions such as the right superior temporal gyrus, rolandic operculum, insula, hippocampus, parahippocampal gyrus, thalamus, lingual gyrus, fusiform gyrus, precentral gyrus, middle frontal gyrus, and the left caudate nucleus, achieved accuracy values ranging from 80% to 88%. Additionally, the f1-score values for these models were found to be within the range of 80% to 89%. Afterward, 3D CNN models constructed using the left supplementary motor area, medial orbital part of superior frontal gyrus, and orbital part of superior frontal gyrus for classifying individuals with BD and individuals with MDD demonstrated accuracy values of 72%, 71%, and 70%, respectively. Correspondingly, the f1-score values for these classifications were observed to be 70%, 71%, and 67%.

Our findings offer a promising approach for integrating rs-fMRI data with 3D CNN techniques for the individual-level classification of affective disorders. Moreover, the observed abnormalities in these brain regions may serve as potential imaging markers for distinguishing patients with MDD and BD from HCs. These biomarkers could also contribute to differentiating individuals with BD from those with MDD. Overall, this study provides a comprehensive understanding of the neurobiology of affective disorders and lays a foundation for developing more precise and personalized diagnostic tools.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Classification and Predictive Modeling ²

Affective Disorders

FUNCTIONAL MRI

Machine Learning

Psychiatric Disorders

Other - deep learning