Introduction

Major depressive disorder (MDD), bipolar disorder (BD), and schizophrenia (SZ) are highly complex psychiatric disorders for which the diagnosis primarily depends on the patient’s clinical symptoms and the psychiatrist’s experiences. Evidence has shown that these mental disorders share clinical symptoms 1, 2, suggesting they may have common pathophysiological mechanisms and it is challenging for the accurate diagnosis of illness. Therefore, there is a push to use computer-aided techniques to help facilitate differential diagnosis and treatment of mental illness.
There has been a recent interest in using deep learning techniques to automatically discriminate individuals with brain disorders from healthy controls (HCs) 3, 4. However, relatively few studies have make classification among mental illness including MDD, BD and SZ. Siamese Network 5 is an artificial neural network and has been used in some small sample multi-class classification studies 6, 7, but it has not been used in the discrimination of psychiatric disorders using neuroimaging measures. The present study aims to identify MDD, BD and SZ based on neuroanatomical features using Siamese Network.

Methods

The participants included four groups: MDD (n = 102), BD (n = 44), SZ (n = 114), and healthy controls (HCs, n = 103). Data were recruited from Henan Mental Hospital between March 2013 and October 2017. The study was approved by the Ethics Committee of Henan Mental Hospital. All the participants provided written informed consent. All the patients had a diagnosis of MDD, BD or SZ schizophrenia confirmed by trained psychiatrists using the Structured Clinical Interview for DSM-IV and were in an acute psychosis state. Exclusion criteria were: other psychiatric disorders; organic causes of depression including heart, liver, or kidney disease; presence of surgically implanted electronic devices. None of the HCs had any personal history of psychotic illness nor any family history of psychosis in their first, second, or third degree relatives. All the participants were Han Chinese in origin, right-handed, and had no contraindications to MRI scanning.
All participants underwent T1-weighted imaging using a 3T Siemens scanner. Acquisition parameters for T1-weighted scans were: repetition time = 2530 ms; echo time (TE) = 2.43 ms; inversion time = 1100 ms; flip angle = 7°; matrix size = 256 × 256 × 192; and voxel size =1×1×1 mm. Foam pads and earplugs were used to reduce head motion and scanner noise.
All T1-weighted images were processed using Statistical Parametric Mapping (SPM12, Wellcome Department of Imaging Neuroscience, London, UK; http://www.fil.ion.ucl.ac.uk/spm/) and Computational Anatomy Toolbox (CAT12). Briefly, the images were bias-corrected and segmented into different tissues including GM, WM, and CSF images. The tissue images were then spatially normalized and resampled to a resolution of 3 × 3 × 3 mm3. To preserve regional volumetric information, the images were modulated by the Jacobian determinants of the deformations during the warping. Voxel-based volumetric features were extracted from the normalized and modulated GM, WM, and CSF images. We noticed that the number of raw features was in the millions. We examined selecting reduced dimensions of features including 2000, 5000, 10000, 20000, according to the order of the feature variance. In the training process of classification, we used the class balanced sampling method to alleviate the problem of sample imbalance. Specifically, we first randomly selected the sample category with equal probability, and then randomly took the sample of the specified category rather than directly taking samples from the whole dataset. We used accuracy, F score, and Macro F as indices for evaluating classification results. For category k, the F score is defined as: F score k = 2×Pk×Rk/(Pk + Rk) Where Pk is accuracy of category k, and Rk is the recall of category k. The Macro F is computed by averaging F scores of all categories. We used an 8-fold cross-validation to test the performance of the Siamese Network and multilayer perception (MLP) network on the dataset.

Results

We tested the performance of the Siamese Network and MLP networks with different numbers of features (N). As shown in Table 1, the Siamese Network achieved improved performance than the MLP network with different features. MDD and SZ had better performance than BD. Finally we achieved a classification accuracy of 46.06% and Macro F1 of 41.47% while ensuring balanced classification.

Conclusion

This study differentiated MDD, BD and SZ from HCs using a Siamese Network that combined GM, WM and CSF volumes. The Siamese Network achieved a better classification performance than the MLP model. Future studies will identify the pattern of structural alterations in patients with psychiatric disorders.

Acknowledgements

This work was supported in part by the Natural Science Foundation of China (Grant No. 31771076).