FRAMEWORK FOR THE CHARACTERIZATION OF HIPPOCAMPUS USING RADIOMIC APPROACH IN FIRST-EPISODE PSYCHOSIS

Abstract Background Volumetric and morphological hippocampal abnormalities are one of the key features of psychotic disorders. Architectural alterations in hippocampal pyramidal neurons have been observed post-mortem in schizophrenia, mainly due to decreased dendritic arborisation and spine density. The in vivo study of grey-matter texture features, which might be sensitive to microstructural abnormalities, may further contribute to the characterisation of hippocampal pathology. However, these characterization techniques have been hardly used in psychiatry. We therefore propose the use of radiomics, able to capture both shape and texture characteristics, for hippocampal characterization in first episode psychosis when compared with healthy volunteers. We evaluated the use of classical statistics and machine learning approaches for differential pattern recognition. Methods For this transversal case-control study, 104 adolescents, 52 with FEP and 52 HV underwent T1-weighted structural MRI in two different scanners: 3T Magnetom Trio-Tim (Siemens, Erlangen, Germany; n=80) and 3T Magnetom Prismafit (Siemens, Erlangen, Germany; n=24). Images were segmented using FreeSurfer v.5.3. and a left-hippocampus mask was used as a ROI for radiomic feature extraction. The Pyradiomics library was used and a total of 100 features, from six feature types were extracted: shape, first order, and other fine texture descriptors. Due to growing concerns about features’ reproducibility and relativeness of intensities, features were extracted multiple times using different yet comparable image preprocessing approaches: normalization within the ROI or across the whole image; and bin width (10, 20, 40) or bin count (100, 50, 15) grey-level discretization. Interclass Correlation Coefficient (ICC(1,3)) was then computed for each of the features and only features with at least moderate ICC (>0.5) were selected. This resulted in the selection of 35 most stable variables, each of which was then extracted from the dataset computed using normalization within the ROI and bin width of 20. Significance of each features was tested between both cohorts using Mann-Whitney test, with α=0.05, False Discovery Rate corrected. Features were also used as inputs to train a Support Vector Classifier (SVC) model with Radial Basis Function (RBF). Accuracy was estimated using 10-fold Cross Validation. Results For the classical statistics evaluation, five features resulted significantly different using Mann-Whitney test: surface volume ratio (p=0.038), kurtosis (p=0.02), grey-level intensity range (p=0.031), skewness (p=0.005) and Imc2 (Informational measure of correlation, p=0.04). However, none of the statistically significant differences survived the FDR-correction. All thirty-five features were also used to train SVC, we selected C=1000 and gamma=0.001 after performing a grid search, and obtained a 68% accuracy with 10-fold cross validation. Discussion The proposed framework constitutes a proof-of-concept approach for the complex hippocampal characterization based on radiomics. Although classical statistical tests were not conclusive, the tendency show that not only shape (volumetric and morphological) but also texture features might provide meaningful information for the characterization of the hippocampus independently. The 68% accuracy is a moderate indicator of discriminative power of the combination of these features, although further analysis should be performed.