Rozpoznávání pacientů s první epizodou schizofrenie s využitím obrazů mozku z MRI
| Title in English | Recognition of the First-Episode Schizophrenia Patients with the Use of MRI Brain Images |
|---|---|
| Authors | |
| Year of publication | 2011 |
| Type | Article in Periodical |
| Magazine / Source | Clinician and Technology (Lékař a technika) |
| MU Faculty or unit | |
| Citation | |
| Field | Neurology, neurosurgery, neurosciences |
| Keywords | image registration, pattern recognition, computational neuroanatomy, schizophrenia |
| Description | Schizophrenia is one of psychiatric disorders which lead often to violation of the social and economic competencies of the patients. The possibility to uncover potentially high-risk patients may enable searching for preventive and therapeutic strategies, which would be able to eliminate or minimize the adverse effects of the disease. Deformation-based morphometry (DBM) has been used to uncover structural inter-group differences in MRI-based neuropsychiatric studies recently. We use 3-D deformation fields resulting from cross-subject registrations to construct classifiers which are able to recognize first-episode schizophrenia patients from healthy controls. The k-Nearest Neighbors (k-NN) and the Support Vector Machines (SVM) classification methods are compared in terms of their sensitivity, specificity, overall accuracy and precision. High-resolution T1-weighted MRI brain images of 173 male subjects were obtained with a Siemens 1.5 T system. The group contained 49 patients with first-episode schizophrenia (FES), and 124 healthy controls (NC). The brain template from Simulated Brain Database was used as the reference anatomy in the high-dimensional nonlinear registration technique based on the maximization of normalized mutual information. The resulting spatial deformations were converted into scalar fields by calculating Jacobian determinants, which characterize local volume shrinkages or enlargements in each voxel. The scalar fields were then analyzed voxel-by-voxel using t-tests with the age of the subjects as a covariate. Significant local volume changes were found with a threshold set to p-value<0.05 corrected by false discovery rate method. The k-NN as well as the SVM classifiers used the significant volume changes as the features in the following classification. Multiple leave-one-out cross-validation was used to evaluate the accuracy of the classification. Due to unequal numbers of subjects in each class, 10 randomly selected subsets of healthy controls were used for evaluation to avoid preference of the NC class. Both classifiers were trained with the same 97 (49+49-1) vectors of 70 845 features. The k-NN classifier gave the best results with k=11 and the cosine metric, median sensitivity was 70%, median specificity was 73% and median accuracy was 72%. The linear SVM classifier gave median sensitivity 80%, median specificity 78% and median accuracy 79%. The accuracy of the algortihm is comparable to state-of-the-art methods for MRI-based schizophrenia classification. |
| Related projects: |