A Stacked Generalization of 3D Orthogonal Deep Learning Convolutional Neural Networks for Improved Detection of White Matter Hyperintensities in 3D FLAIR Images

Qualitative evaluation of WMH detection performance by StackGen-Net. Representative axial, coronal, and sagittal slices from a test subject are shown in the left panel. Manual annotations and predictions from StackGen-Net are overlaid in red. A, The insets from coronal images are zoomed in for better comparison of the prediction with the ground truth. Compared with manual annotation, StackGen-Net slightly overestimates the lesion contour. B, A comparison of WMH predictions from the orthogonal CNNs (axial, sagittal, and coronal) is shown. The yellow arrows show WMHs that were missed by a majority of the CNNs in the ensemble. These lesions would have been missed by a simple averaging or majority voting of the orthogonal CNN predictions but are identified correctly by StackGen-Net. Sag indicates sagittal; Cor, coronal.

Boxplot comparison of Dice scores, lesion-based F1 (F1-L), volume difference (VD), and area under precision-recall curve (AUC) scores on the test set. We found a significant improvement in Dice scores, AUC, and F1-L in StackGen-Net compared with other WMH segmentation techniques compared here. The asterisk denotes P < .001 (2-sided paired t test, n = 29).