PREFRONTAL DYSFUNCTION IN PARKINSON`S DISEASE
Transcript
PREFRONTAL DYSFUNCTION IN PARKINSON`S DISEASE
Christian Salvatore First Year Seminar (2012/2013) – XXVIII cycle. Dottorato di Ricerca in Fisica, Scuola di Dottorato di Scienze, Università degli Studi di Milano - Bicocca. Supervisor: Prof. Marco Paganoni External supervisor: Dott.ssa Isabella Castiglioni (IBFM-CNR; HSR) Title Development of a Decision Support System for automatic diagnosis of medical images from MRI studies Background Machine learning methods for the assisted medical diagnosis are recently growing in popularity within neuroimaging community. Supervised machine learning methods are based on algorithms able to automatically extract multiple information from image sets without requiring a-priori hypotheses of where this information may be coded in the images. These methods have been proposed as a revolutionary approach for identifying sensitive biomarkers (or a combination between them) allowing for automatic discrimination of individual subjects. To date, Parkinson's disease (PD) is the second most common neurodegenerative disease affecting millions of people worldwide. One of the most critical issues in the clinical practice of PD is to achieve a definite individual differential diagnosis. This is due to 1) poor accuracy, specificity and sensitivity at a visual inspection and 2) difficulty in differentiation of PD from other parkinsonian conditions, such as Progressive Supranuclear Palsy (PSP). Effective and accurate diagnosis of PD and PSP by means of MRI biomarkers has recently attracted strong attention. So far, several biomarkers have been shown to be sensitive to the diagnosis of PD with respect to PSP. However, to date, the only validated MRI-related measurement employed in clinical practice of PD is based on manual morphometric quantification of conventional MR images. Purpose The aim of this first year of work was to implement and optimize a supervised machine-learning method able to perform automatic individual differential diagnosis of PD and PSP by means of structural T1-weighted Magnetic Resonance Images (MRIs). This method was based on Principal Component Analysis (PCA) as feature extraction technique and on Support Vector Machines (SVMs) as classification algorithm. In order to allow the identification of new MRI-related biomarkers useful for the diagnosis of PD and PSP, I also generated image maps of pattern distribution of brain structural differences, which reflect the importance of each image voxel for the SVM classification. Materials and Methods In this study, MR images of 56 patients and 28 healthy control subjects were acquired. The group of 56 patients consisted of 28 patients with clinically diagnosed PD and 28 patients with clinical diagnosis of probable or possible PSP. Bain structural MRI studies were performed by a 1.5-T unit (Signa NV/I; GE Medical Systems, USA). We acquired a T1weighted 3D dataset for each subject. Original images were cropped, re-oriented and converted from DICOM format to 3D NIfTI format using the dcm2nii tool included in the MRICron software. Skull-stripping of the resulting volumes was made using the bet tool of the FSL 4.1 software. Each volume was normalized to MNI space by co-registration to the MNI template (MNI152_T1_1mm_brain) included in the FSL 4.1 software. All images were then imported in Matlab. The final brain volume of MR images consisted of 145x178x133 voxels. A machine-learning method able to perform individual differential diagnosis of PD and PSP by means of structural T1-weighted Magnetic Resonance Images (MRI) was implemented: this method was based on Principal Components Analysis (PCA) as feature extraction technique 1 and on Support Vector Machines (SVMs) as classification algorithm. PCA transformation was applied on the images in order to reduce the dimension of the data space without losing relevant information. The SVM classifier was trained using the resultant principal components and the associated labels as features for the SVM training. To improve generalization, we used a linear kernel. Accuracy, specificity and sensitivity of the SVM classifier were assessed by a leave-one-out (LOO) cross-validation strategy. We also adopted another validation method using N/2 randomly chosen subjects for training the classifier and the remaining N/2 subjects for testing. Maps of voxel-based pattern distribution of brain structural differences were generated, reflecting the importance of each voxel for the SVM classification. Results The classifier allowed individual differential diagnosis with the following accuracy (specificity/sensitivity): in LOO validation, PD vs. Controls 92.7 (92.3/93.4)%, PSP vs. Controls 97.0 (98.2/95.9)%, PSP vs. PD 98.2 (98.8/97.8)%; in N/2 validation, PD vs. Controls 93.5 (90.6/97.4)%, PSP vs. Controls 92.2 (92.5/92.4)%, PSP vs. PD 92.2 (91.3/94.4)%. The following MRI-related brain biomarkers were identified to be used for differential diagnosis of PD and PSP: midbrain, pons, corpus callosum and thalamus, four critical regions which are highly consistent with typical neuropathological and imaging findings described in patients with PSP. Conclusions In this first year of work I implemented and optimized a voxel-based pattern recognition method for the differential diagnosis of PD and PSP. This method was applied to 3D T1-weighted structural neuroimaging data, returning Accuracy, Sensitivity and Specificity values > 90%, higher than other published methods. Moreover, I generated maps of pattern distribution of brain structural differences, thus allowing the identification of MRI-related biomarkers for differential diagnosis of PD and PSP. These findings encourage the use of computer-based SVM analysis in clinical neuroradiology. Future Developments Next steps will include the implementation of a classifier able to return probability outputs. New imaging datasets will be acquired. The classifier will be validated using these new datasets. Attended Schools, Courses and Seminars Schools INIT/AERFAI International Summer School on Machine Learning. Benicàsim, Spain. June 2428, 2013 [Seminar: September, 19th 2013]. Courses Analisi Statistica dei Dati, M. Bonesini [Exam: October, 2013]. Connecting with English, J. Weekes, February – December, 2013. Seminars Outlook in Neutrino Physics. Pilar Hernandez. Università degli Studi di Milano-Bicocca. February, 13th 2013. LHC And The New Technologies To Go Beyond The Higgs Boson. Lucio Rossi. Università degli Studi di Milano-Bicocca. March, 13th 2013. 2 The Origin Of Life: From Geophysics To Biology. Development Of An Artificial Cell From SelfOrganization To Replication. Albert Libchaber. Università degli Studi di Milano-Bicocca. April, 10th 2013. The Era Of Cosmic Reionization. Francesco Haardt. Università degli Studi di Milano-Bicocca. May, 15th 2013. Publications (2012/2013) Indexed International Papers Grosso, E., López, M., Salvatore, C., Gallivanone, F., Di Grigoli, G., Valtorta, S., Moresco, R.M., Gilardi, M.C., Ramírez, J., Górriz, J.M. and Castiglioni, I. (2012). A decision Support System for the assisted diagnosis of brain tumors: a feasibility study for 18F-FDG PET preclinical studies. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS: pp.6255-6258; ISSN: 1557170X; ISBN: 978-142444119-8; DOI: 10.1109/EMBC.2012.6347424. Gallivanone, F., Di Grigoli, G., Salvatore, C., Valtorta, S., Gilardi, M.C., Moresco, R.M. and Castiglioni, I. (2012). Acute stress studies in rats by 18F-FDG PET and SPM. IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC) M15-4: pp.2886-2889; ISBN: 978-1-4673-2029-0. Castiglioni, I., Cerasa, A., Salvatore, C., Gallivanone, F., Augimeri, A., Lopez,M., Gilardi, M.C. and Quattrone, A. (2013). Machine Learning Performs Differential Individual Diagnosis of PD and PSP By Brain MRI Studies. Movement Disorders. 28 Suppl 1:180. Cava, C., Zoppis, I., Mauri, G., Ripamonti, M., Gallivanone, F., Salvatore, C., Gilardi, M.C. and Castiglioni, I. (2013). Combination of gene expression and genome copy number alteration has a prognostic valure for breast cancer. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS. Salvatore, C., Cerasa, A., Castiglioni, I., Gallivanone, F., Augimeri, A., Lopez, M., Arabia, G., Morelli, M., Gilardi, M.C. and Quattrone, A. (2013). Machine Learning Performs Differential Individual Diagnosis of Parkinson's Disease and Progressive Supranuclear Palsy by Brain MRI Studies. J Neurosci Methods [under minor revision]. International Book Chapters Cava, C., Gallivanone, F., Salvatore, C., Della Rosa, P. and Castiglioni, I. (2013). Bioinformatics clouds for high-throughput technologies. Cloud Infrastructures for Big Data Anaytics, IGI Global [accepted] Gallivanone, F., Salvatore, C., Cava, C., Della Rosa, P. and Castiglioni, I. (2013). Applications of image processing for classification of neuroimages for early and differential diagnosis of dementia and brain cancer diseases. Global Trends in Knowledge Representation and Computational Intelligence, IGI Global [accepted]. Salvatore, C., Gallivanone, F., Della Rosa, P., Cava, C. and Castiglioni, I. (2013). Multiple Classifier systems for classification of medical images. Image Analysis using Computational Intelligence [submitted]. 3 International Conference Proceedings Gallivanone, F., Di Grigoli, G., Salvatore, C., Valtorta, S., Grosso, E., Gilardi, M.C., Moresco, R.M. and Castiglioni, I. (2012). SPM for activation studies in rats on stress conditions. 6th Hot Topics in Molecular Imaging Conference (TOPIM): p.44. Gallivanone, F., Di Grigoli, G., Salvatore, C., Belloli, S., Valtorta, S., Raccagni, I., Gilardi, M.C., Castiglioni, I. and Moresco, R.M. (2013). Feasibility of supervised machine learning technique for assisted diagnosis of cancer: application to PET studies in small animals. European Molecular Imaging Meeting (EMIM), 8th annual meeting of the European Society for Molecular Imaging (ESMI), Turin, Italy. National Conference Proceedings Gallivanone, F., Grosso, E., Di Grigoli, G., Salvatore, C., Valtorta, S., Gilardi, M.C., Moresco, R.M. and Castiglioni, I. (2012). Statistical Parametric Mapping for activation studies in rats. Atti del Congresso Nazionale di Bioingegneria: p.147; ISBN: 978 88 555 3182-5147. 4