Image Bank

Retrospective examinations were collected from a 3.0 Tesla MRI scanner in Botucatu and Campinas Medical School Hospital (UNESP and UNICAMP – Brazil). Patients were selected according to the following criteria:

• Inclusion criteria: patients who underwent MRI exams between 2010 and 2017 with diagnosis of inflammation or tumors. All selected patients had their diagnosis confirmed by histopathological or CSF analysis examinations and follow-up exams.
• Exclusion criteria: patients who had brain biopsy or surgery before the MRI acquisition, lesions smaller than 10mm, and patients with intracranial malformations. 

The selection resulted in an image database with 156 MRI examinations in four different MRI sequences, T1- and T2-weighted images, diffusion, and Fluid Attenuated Inversion Recovery (FLAIR). The complete list with all pathologies that were selected for this study, with the number of patients, and their mean lesion size is shown in Table 1. 

Feature extraction 

In the first step of this approach, a group of experienced radiologists analyzed all MRI sequences, in axial orientation: T1, T1 with contrast, T2 and FLAIR. Images had different sizes and were all resized to 240 x 240 pixels. The radiologists selected the slice in which the lesion appeared with its largest diameter. Regions of Interests (ROI) were positioned within each lesion. An example of ROI positioning is presented in Fig. 1

The number of ROIs varied in each slice according to each lesion size. Then, the features: mean, standard deviation, entropy, kurtosis, skewness and correlation were extracted from those ROIs and textural features were extracted using GLCM [17], GLRL [18], and Wavelet’s Transform methods. 

These firsts steps were performed using Matlab software R2017a. The features extracted were processed in the software Orange Canvas® for application of ML methods that performed the classification.

Machine learning classification

Machine learning (ML) approaches were performed using Orange Canvas® software. In order to determinate the best approach, we compared five types of image: T1, T1 with intravenous contrast (T1 C+), T2, diffusion and FLAIR. To each image, we input the data in three ways: all extracted features as raw data; the mean of all ROIs per patient to each extracted features; and only the five best-ranked features from raw data. To rank the features, we used Information Gain Ratio and Gini Decrease scoring methods.

We used three different methods of ML: SVM, kNN, and Random Forest. All three methods used the textural features to train and then test data collected with the GLCM, GLRL, and Wavelet methods.

We used the parameters area under the curve (AUC), accuracy (CA), F-score (F1), Precision and Recall to determinate how efficiently the model classified our groups.