Recognition of First-episode Schizophrenia from MRI Data with the Use of Artificial Neural Networks
-
1
Masaryk University, Institute of Biostatistics and Analyses, Czechia
-
2
Masaryk University, Research Centre for Toxic Compounds in the Environment, Czechia
-
3
Masaryk University and University Hospital Brno, Department of Psychiatry, Czechia
INTRODUCTION
Schizophrenia as a severe disabling psychiatry disorder. If the neuroscientific community invented an algorithm that could find the information about first-episode schizophrenia (FES) in the brain images, it could help to objectively establish the diagnosis as soon as the patient’s brain is scanned in the particular imaging device. Early diagnosis would help to tackle the symptoms threatening the patients and their surroundings in the early stage of the disease by deployment of antipsychotics. Since modern computers provide sufficient power to calculate even computationally expensive tasks, self-adaptive models such as neural networks get into the foreground. We explore three types of neural networks – multilayer perceptron (MLP), radial basis function network (RBF), learning vector quantization network (LVQ) initialized using Kohonen’s self-organizing map [1] – for a classification of first-episode schizophrenia based on structural MRI.
METHODS
The dataset used consisted of 52 patients and 52 healthy control subjects structural MR images acquired on 1.5 T magnetic resonance imaging device Siemens at University Hospital Brno. All the images were preprocessed using optimized voxel-based morphometry in several steps [2]: correction for bias-field inhomogeneity, spatial normalization and segmentation, modulation and Gaussian smoothing. Only the gray matter densities were extracted and used to adapt the models. To further reduce the data dimensionality, two-sample t-tests were applied and only the several lengths of the most significant voxels that carry the information about grey matter densities were used to construct feature vectors.
The extracted and selected features provided the information for the adaptation of the artificial neural networks (ANNs) and support vector machines (SVM) that were used as a reference method. The power of all the network types is directly affected by their parameters’ settings. Although the setting of weights is optimized during the training phase, other parameters such as the number of input and hidden neurons, the spread of RBF neurons and LVQ learning algorithms must be specified by the modeler. We explored several of these parameters and observed how they can affect the classifiers’ performance measures – overall accuracy (OA), sensitivity (Sen) and specificity (Spe).
RESULTS
As expected, the settings of the explored parameters matter. The most successful architectures among the investigated range of parameter settings of each neural network type were as follows: MLP which consisted of 700 input neurons and 10 hidden neurons had OA 0.70 (Sen 0.68, Spe 0.72), RBF network with 600 input neurons, 10 hidden neurons and spread equal to 3 reached OA 0.76 (Sen 0.77, Spe 0.75), LVQ network with 700 input and 10 hidden neurons adapted by LVQ 1 learning algorithm achieved OA 0.67 (Sen 0.72, Spe 0.63) and by LVQ2.1 learning algorithm achieved OA 0.69 (Sen 0.73, Spe 0.65). Since SVM reached only OA 0.64 (Sen 0.56, Spe 0.71), all the ANN types improved diagnosis of FES compared to this commonly used method. Although the success of the classifiers differs, McNemar’s test revealed significant difference only between RBF network and SVM (p = 0.02) in behalf of the former one.
CONCLUSION
Artificial neural networks are tools that can keep up with or can be even better than the traditional method for pattern recognition which is SVM and therefore, they can help with computer-aided schizophrenia diagnosis. The accuracy required for potential clinical practice in psychiatry has not been reached here though and hence our further work will focus on ensemble learning and deep learning methods that can potentially improve the classification performance.
References
[1] T. Kohonen, Self-organizing maps, 3rd ed. Berlin ; New York: Springer, 2001.
[2] C. D. Good, I. S. Johnsrude, J. Ashburner, R. N. Henson, K. J. Friston, and R. S. Frackowiak, “A voxel-based morphometric study of ageing in 465 normal adult human brains,” Neuroimage, vol. 14, no. 1 Pt 1, pp. 21–36, Jul. 2001.
Keywords:
MRI,
Schizophrenia,
artificial neural networks,
Classification,
Computer-aided diagnostics
Conference:
SAN2016 Meeting, Corfu, Greece, 6 Oct - 9 Oct, 2016.
Presentation Type:
Poster Presentation in SAN2016 Conference
Topic:
Posters
Citation:
Vyskovsky
R,
Schwarz
D,
Janousova
E and
Kasparek
T
(2016). Recognition of First-episode Schizophrenia from MRI Data with the Use of Artificial Neural Networks.
Conference Abstract:
SAN2016 Meeting.
doi: 10.3389/conf.fnhum.2016.220.00073
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
29 Jul 2016;
Published Online:
01 Aug 2016.
*
Correspondence:
Prof. Daniel Schwarz, Masaryk University, Institute of Biostatistics and Analyses, Brno, Czechia, schwarz@iba.muni.cz