255. Automated classification of Resting State fMRI networks using Machine Learning algorithms. (December 2018)

• Record Type: Journal Article
• Authors: Pinardi, C.; Ortenzia, O.; Gardini, S.; Aldigeri, R.; Micheli, M.; Spigoni, V.; De Cais, A.; Ghetti, C.
• Is Part Of: Physica medica. Volume 56(2018)Supplement 2
• Page Start: 219
• Page End: 220
• Publication Date: 2018-12
• Subjects: Medical physics -- Periodicals; Biophysics -- Periodicals; Biophysics -- Periodicals; Imagerie médicale -- Périodiques; Radiothérapie -- Périodiques; Rayons X -- Sécurité -- Mesures -- Périodiques; Physique -- Périodiques; Médecine -- Périodiques; 610.153
• DOI: 10.1016/j.ejmp.2018.04.266
• Languages: English
• ISSNs: 1120-1797

Abstract

Abstract : Purpose: The need of an automatic reorganization of the massive Resting State (RS) output is impelling due to the spreading of the RS fMRI[1], as a huge number of RS network is carried out for each subject and only a few of them are interesting. The aim of the study is to use a Machine Learning algorithm to identify four relevant networks (Default Mode, Auditory, Visual, Frontal), developing an automated protocol of network classification. Methods: RS analysis has been carried out with CONN software[2] in 21 elderly subjects (11 M–10 F, 73 ± 4 years), acquired in GE 3T scanner with a EPI sequence (TR/TE 2000/30 ms, 3.2 × 3.2 × 3.4 mm 3, Volumes 250): the raw fMRI images were spatially and temporally preprocessed with standard analysis steps and subsequently a Independent Component Analysis (ICA) was performed to carry out the RS networks for each subject. The four classes of relevant networks were tested against a class of random chosen networks in PRoNTo software[3] with a multiclass Gaussian Process. After the cross-validation test phase, the weight of single voxels into the classification process was obtained for each network class. Results: PRoNTo successfully classified the networks with a permutation test obtaining a total accuracy of 92.38% (p < 0.002). The results of class accuracy and predictive values are shown inTable 1 . Fig. 1 represents the confusion matrix. Conclusion: The Machine Learning algorithm correctly identified four relevant RS networks against random ones, showing the possibility to automatize a protocol of RS network classification. Funding: DRINN PROJECT – Diabetes Research Innovation 2015.