Purpose: We developed a functional magnetic resonance imaging (fMRI)-based radiomics model for the diagnosis of major depressive disorder (MDD).
Methods: A total of 146 participants (84 healthy controls, 62 MDD patients) were randomly divided into a training set (n = 110) and a validation set (n = 36) using stratified sampling. The resting-state fMRI images were preprocessed, including co-registration, correction, normalization and segmentation, using public toolbox DPABI in MATLAB, resulting in three amplitude of low-frequency fluctuation (ALFF) images, three fractional ALFF (fALFF) images, and four regional homogeneity (ReHo) images for each participant. We extracted 1118 radiomics features from each preprocessed images using an open-source python package Pyradiomics, including first-order statistical descriptors, high-order texture features, wavelet features, and Laplacian of Gaussian (LoG) features. The Shapiro-wilk test was used to calculate the normality of data. The differences between healthy controls and MDD patients in normally distributed features were calculated using two-sample t-test, otherwise, two-sided Wilcoxon rank sum test. All statistical analyses were performed by using SPSS software. The features with P < 0.05 were enrolled into least absolute shrinkage and selection operator (LASSO) regression model for further feature selection. The most valuable features selected by LASSO were used for a support vector machine (SVM) model.
Results: Among the ten models that were built based on different fMRI images processing methods, three models performed best in classifying healthy controls and MDD patients, including ALFF, mfALFF (i.e., each voxel’s fALFF was divided by the global mean fALFF), and ReHo models. The ALFF model outperformed mfALFF and ReHo (accuracy: 86% and 83%, area under curve [AUC]: 0.88 and 0.88, sensitivity: 75% and 75%, specificity: 92% and 87.5%) models with an accuracy of 100% (AUC: 1).
Conclusion: This study demonstrates that the fMRI-based radiomics model can provide favorable predictive efficacy for the diagnosis of MDD.
Funding Support, Disclosures, and Conflict of Interest: This study was supported by Taishan Scholars Program of Shandong Province (No. ts201712065) and Academic promotion programme of Shandong First Medical University (2019QL009)