Abstract: Introduction: Deep learning (DL) models are powerful multivariate models having great potential for psychiatric disorder research [1], where brain markers have yet to be identified. A promising DL approach is based on anomaly detection using normative autoencoder (AE) models [2], in which psychiatric patients could be detected as deviating samples based on brain features. Aim: The primary study aim was to develop a generalizable and multi-site framework for AE-based discrimination of psychiatric disorders based on structural magnetic resonance imaging (sMRI) data. The desired pipeline should be able to (i) integrate confounding effect correction into internal and external validation frameworks so to extend the application to data from new sites, (ii) automatically discriminate patients from healthy controls (HC) through an anomaly detection strategy. Methods: T1-weighted sMRI scans were acquired in six sites using 3T scanners (training set, n=460 HC) and in one site using a 1.5T scanner (external set, n=65 HC,48 schizophrenics (SCZ), 20 major depression (MDD) and 43 bipolar disorder (BD)), all sites were part of the StratiBip consortium. All images were pre-processed with the SPM CAT12 toolbox. Regional grey matter (GM) morphological features were extracted (68 cortical thicknesses and 52 GM volumes). The multisite features were harmonized using ComBat tool [3] and corrected for age, sex, and harmonized total intracranial volume using linear regression. The parameters were estimated in the training set and applied to both training and external sets [4]. A different harmonization strategy had to be designed for the external set, as there were no training set ComBat parameters estimations for this ?new? site. Using ComBat, we used the harmonized training set as reference to estimate the HC external set site parameters [5]. The model underwent a 10-fold cross-validation for hyperparameter tunning where the data corrections were performed within each fold. After the AE was trained with the training set, we extracted the subjects' reconstruction error for the external set and evaluate model performance in terms of AUC-ROC, using the mean square reconstruction error (MSE) of each subject and respective labels to discriminate HC vs. patients. Results: We evaluated the harmonization quality by looking at the 2 principal components and verified that data clusters associated with site were no longer distinguished after harmonization procedure. The normative model performance was, for training set, MSE=0.0262 ±0.0001, and HC external set, MSE=0.0690±0.0004. The discrimination results for the external set were respectively, HC vs. SCZ (AUC=0.73), HC vs. MDD (AUC=0.60), HC vs. BD (AUC=0.70). Conclusions: This study proposes a generalizable and flexible DL approach that enables the usage of multi-site datasets for neuroanatomic biomarker research for psychiatric disorders. Our hypothesis that psychiatric disorder patients could be discriminated from HC as deviating samples was verified for the external set used. These results pave the way for the use of this approach for the extraction of abnormal brain feature patterns being shared and distinct between affective and psychotic disorders.

Autoría

Descargar referencia