Identifying Depressive Disorder: Perspectives from an Individualized Radiomics-Based Structural Similarity Network

Motivation: Lacking a definitive standard for diagnosis, major depressive disorder (MDD) is currently difficult to accurately identify and manage effectively. Goal(s): This study aimed to explore the potential of the individualized radiomics-based structural similarity network (iRSSN) for identifying MDD. Approach: Subject-specific iRSSN matrices for 1,029 MDD patients and 927 healthy controls were calculated as Pearson correlation coefficients of 43 radiomic features between each pairwise of 3D-T1w ROIs based on the Brainnetome atlas. An iRSSN-based classification model was developed using logistic regression with Lasso feature reduction and assessed via five-fold cross-validation. Results: The accuracy the classifier achieved was 80.6% (95%CI:77.8-83.5%, AUC:0.891, sensitivity:0.80, specificity:0.81, F1 Score:0.80). Impact: Individualized radiomics-based structural similarity network shows great potential for serving as more reliable features in MDD diagnosis.