Reliability of an automatic classifier for brain enlarged perivascular spaces burden and comparison with human performance

In the brain, enlarged perivascular spaces (PVS) relate to cerebral small vessel disease (SVD), poor cognition, inflammation and hypertension. We propose a fully automatic scheme that uses a support vector machine (SVM) to classify the burden of PVS in the basal ganglia (BG) region as low or high. We assess the performance of three different types of descriptors extracted from the BG region in T2-weighted MRI images: (i) statistics obtained from Wavelet transform's coefficients, (ii) local binary patterns and (iii) bag of visual words (BoW) based descriptors characterizing local keypoints obtained from a dense grid with the scale-invariant feature transform (SIFT) characteristics. When the latter were used, the SVM classifier achieved the best accuracy (81.16%). The output from the classifier using the BoW descriptors was compared with visual ratings done by an experienced neuroradiologist (Observer 1) and by a trained image analyst (Observer 2). The agreement and cross-correlation between the classifier and Observer 2 (kappa = 0.67 (0.58-0.76)) were slightly higher than between the classifier and Observer 1 (kappa = 0.62 (0.53-0.72)) and comparable between both the observers (kappa = 0.68 (0.61-0.75)). Finally, three logistic regression models using clinical variables as independent variable and each of the PVS ratings as dependent variable were built to assess how clinically meaningful were the predictions of the classifier. The goodness-of-fit of the model for the classifier was good (area under the curve (AUC) values: 0.93 (model 1), 0.90 (model 2) and 0.92 (model 3)) and slightly better (i.e. AUC values: 0.02 units higher) than that of the model for Observer 2. These results suggest that, although it can be improved, an automatic classifier to assess PVS burden from brain MRI can provide clinically meaningful results close to those from a trained observer.