Tree-Based Ensemble Learning Techniques in the Analysis of Parkinsonian Syndromes

• Published in: Medical Image Understanding and Analysis Vol. 723; pp. 459 - 469
• Main Authors: Górriz, J. M., Ramírez, J., Moreno-Caballero, M., Martinez-Murcia, F. J., Ortiz, A., Illán, I. A., Segovia, F., Salas-González, D., Gomez-Rio, M.
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2017; Springer International Publishing
• Series: Communications in Computer and Information Science
• Subjects: Atypical parkinsonism; Bagging; Boosting; Decision tress; Ensemble pattern recognition; Single photon emission computed tomography (SPECT)
• ISBN: 9783319609638; 3319609637
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-3-319-60964-5_40

$$^{123}$$ I-ioflupane single photon emission computed tomography (SPECT) is a standard and well-known imaging modality in the medical practice for the diagnosis of Parkinson’s disease (PD). That said, atypical parkinsonian syndrome (APS), a symptom-related disease to PD, detection is yet considered inconsistent at least based on visual inspection on region of interests (ROIs). Although some machine learning approaches have been proposed in this regard, in this paper we take up this matter again by applying advanced image processing techniques based on ensemble learning in order to discriminate PD from the various APS, included in the group denominated as P plus. This study enrolled 168 subjects including followed-up patients with degenerative parkinsonism and normal controls undergoing $$^{123}$$ I-ioflupane SPECT at the “Virgen de las Nieves” Hospital, Spain in the last years: 45 Normal, 75 PD, 31 APs (including multiple system atrophy (MSA), progressive supranuclear palsy (PSP) and corticobasal degeneration (CBD) patients) and 17 controls. Several advanced ensemble techniques and feature extraction methods were applied voxel-wise to the analysis of the SPECT images using robust classifiers based on decision trees. The system is trained by means of boosting and bagging algorithms and their performance control is specified in terms of the classification error and the received operating characteristics curve (ROC) using 10-fold cross validation. By the use of these statistical validation methods it was possible to confirm that this modality may be useful for discriminating the abnormal patterns under study.