Research Article: Cerebral 18F-FDG PET in macrophagic myofasciitis: An individual SVM-based approach

Date Published: July 13, 2017

Publisher: Public Library of Science

Author(s): Paul Blanc-Durand, Axel Van Der Gucht, Eric Guedj, Mukedaisi Abulizi, Mehdi Aoun-Sebaiti, Lionel Lerman, Antoine Verger, François-Jérôme Authier, Emmanuel Itti, Bin Liu.

http://doi.org/10.1371/journal.pone.0181152

Abstract

Macrophagic myofasciitis (MMF) is an emerging condition with highly specific myopathological alterations. A peculiar spatial pattern of a cerebral glucose hypometabolism involving occipito-temporal cortex and cerebellum have been reported in patients with MMF; however, the full pattern is not systematically present in routine interpretation of scans, and with varying degrees of severity depending on the cognitive profile of patients. Aim was to generate and evaluate a support vector machine (SVM) procedure to classify patients between healthy or MMF 18F-FDG brain profiles.

18F-FDG PET brain images of 119 patients with MMF and 64 healthy subjects were retrospectively analyzed. The whole-population was divided into two groups; a training set (100 MMF, 44 healthy subjects) and a testing set (19 MMF, 20 healthy subjects). Dimensionality reduction was performed using a t-map from statistical parametric mapping (SPM) and a SVM with a linear kernel was trained on the training set. To evaluate the performance of the SVM classifier, values of sensitivity (Se), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV) and accuracy (Acc) were calculated.

The SPM12 analysis on the training set exhibited the already reported hypometabolism pattern involving occipito-temporal and fronto-parietal cortices, limbic system and cerebellum. The SVM procedure, based on the t-test mask generated from the training set, correctly classified MMF patients of the testing set with following Se, Sp, PPV, NPV and Acc: 89%, 85%, 85%, 89%, and 87%.

We developed an original and individual approach including a SVM to classify patients between healthy or MMF metabolic brain profiles using 18F-FDG-PET. Machine learning algorithms are promising for computer-aided diagnosis but will need further validation in prospective cohorts.

Partial Text

Macrophagic myofasciitis (MMF) is an emerging condition with highly specific myopathological alterations found at deltoid muscle biopsy assessing persistence of aluminum hydroxide adjuvant particles within macrophages that may occur following intramuscular vaccine injections (#ORPHA592, http://www.orpha.net) [1,2]. In most patients, clinical manifestations typically associated with MMF include arthromyalgias, chronic fatigue and cognitive impairment, occurring several months or years after the last vaccine injection [3–5]. Few functional SPECT and PET studies have investigated this cognitive disorder [6–8]. A peculiar spatial pattern of a cerebral glucose hypometabolism involving occipito-temporal cortex and cerebellum have been reported [7]; however, the full pattern is not systematically present in routine interpretation of scans, and with varying degrees of severity depending on the cognitive profile of patients [8]. Thus, the identification of an individual biomarker is needed.

The aim of the current study was to generate and evaluate a support vector machine procedure to classify patients between healthy or MMF 18F-FDG brain profiles. The generated t-test mask was that of our previous findings [8], corresponding to a symmetrical pattern of hypometabolism involving occipital, temporal lobes but also the limbic system, fronto-paritetal cortex and both hemispheric cerebellum. Nevertheless, these analyses were performed at the scale of a group where the signal was smoothed. An ANCOVA was performed between the two groups of healthy or MMF patients which aims to compare the means of each voxels, and how they statistically differ with additional qualitative covariate (age). Here, we present an original procedure where we build a SVM classifier at the patient level to classify each patient between healthy or pathologic brain profiles. Other classifiers could have been used such as artificial neural networks, random forest classifiers or Bayesian techniques. Artificial neural networks or multilayer perceptron have the ability to learn nonlinear models by back-propagating errors [19]. However, they require bigger populations to train and have non-convex cost function that may not always converge to the same values. Similarly, random forest classifiers (which are random tree classifiers based on bootstrapping) may not give identical results each time the algorithm is run. Inversely, the SVM approach has the advantage to be easily implemented, fast to train and works well when the number of parameters is higher than the number of patients.

In conclusion, the SVM procedure appears as a useful individual tool to classify subjects between healthy or MMF 18F-FDG PET brain profiles. We developed an original approach with first a features reduction using Z-score maps with SPM analysis. Secondly, an SVM procedure was trained and coefficients weights were studied as they seem linked to the Z-score value. Machine learning algorithms are promising for computer-aided diagnosis either for classification or regression problematics. As they will gain more and more importance in the year to come, they will challenge traditional interpretation of examination but surely will help physicians and healthcare providers more comprehensive insights on diverse pathologies.

 

Source:

http://doi.org/10.1371/journal.pone.0181152

 

0 0 vote
Article Rating
Subscribe
Notify of
guest
0 Comments
Inline Feedbacks
View all comments