Automated Detection of Lupus White Matter Lesions in MRI

Abstract

Brain magnetic resonance imaging provides detailed information which can be used to detect and segment white matter lesions (WML). In this work we propose an approach to automatically segment WML in Lupus patients by using T1w and fluid-attenuated inversion recovery (FLAIR) images. Lupus WML appear as small focal abnormal tissue observed as hyperintensities in the FLAIR images. The quantification of these WML is a key factor for the stratification of lupus patients and therefore both lesion detection and segmentation play an important role. In our approach, the T1w image is first used to classify the three main tissues of the brain, white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF), while the FLAIR image is then used to detect focal WML as outliers of its GM intensity distribution. A set of post-processing steps based on lesion size, tissue neighborhood, and location are used to refine the lesion candidates. The proposal is evaluated on 20 patients, presenting qualitative, and quantitative results in terms of precision and sensitivity of lesion detection [True Positive Rate (62%) and Positive Prediction Value (80%), respectively] as well as segmentation accuracy [Dice Similarity Coefficient (72%)]. Obtained results illustrate the validity of the approach to automatically detect and segment lupus lesions. Besides, our approach is publicly available as a SPM8/12 toolbox extension with a simple parameter configuration.

Keywords: automatic lesion detection and segmentation; image analysis; lupus disease; magnetic resonance images.

Figure 1

Scheme of the full pipeline. T1w and FLAIR images are the original subject images, which require a co-registration. Once co-registered they undergone a separate pre-processing, T1w by SPM8/12 to obtain the tissue segmentation and the brain mask, while FLAIR is denoised and unbiased by separate methods. The atlas and its structures also belong to the SPM space, so they can be brought to the subject space as the tissue segmentation and the brain mask. The corrected FLAIR image, the tissue segmentation and the posterior fossa mask are the inputs of the WML segmentation tool.

Figure 2

Neighboring rule. FLAIR 2D axial slice (A) showing 2 lesions (in green) of 4 and 5 voxels, respectively (B), both completely surrounded by white matter. Original image and tissue segmentation result of two slices forward are shown in (C,D). The bottom row shows four zooms of the original image, ground truth (green), candidates regions (red), and final lesion segmentation (blue).

Figure 3

Joint evaluation of both α and λ parameters (x and y axes, respectively) for each training set within the two-fold cross-validation. Each position in the map represents the mean F-score for a specific tissue neighborhood ratio λ and the initial candidate lesions adjustment α. Reddish colors show higher mean F-score values.

Figure 4

Bar plots of each patient representing the DSC, TPR, and PPV values. The population is stratified in three groups depending on the GT number of lesions, from left to right: < 5; [5−25]; >25 lesions per patient.

Figure 5

Correlation with number of lesions (stratified by the three groups) on the left and lesion volume, in terms of voxels, on the right.

Figure 6

Bland-Altman plot of the lesion volume for those cases below 500 mm³ of average volume. There is also one outlier not represented with 400 mm³ of difference volume and 370 mm³ of average volume. Besides, two scatter points over 500 mm³ of average volume have also been removed, although the difference volume was 10 mm³ for both. The* represents the mean difference volume when removing the outlier from the analysis.

Figure 7

Qualitative results of the approach. First row of each patient shows the original FLAIR image and second row shows the automatic segmentation (green, TP; red, FP; and yellow, FN).