Convolutional Neural Network for Automated FLAIR Lesion Segmentation on Clinical Brain MR Imaging

Schematic of the CNN U-net architecture. The architecture uses a 3D region-based approach for training and validation. The sample MR FLAIR images are from a patient with progressive multifocal leukoencephalopathy. Max indicates maximum.

Representative slices from validation samples of FLAIR MR brain images (A) with CNN-based (B) and manual lesion segmentations (C), with predicted or ground truth lesion segmentations overlaid in red. The CNN performs well on a variety of different neurologic disorders, here shown in cases of multiple sclerosis, SVID, low grade-glioma, primary CNS lymphoma, adrenoleukodystrophy, and toxic leukoencephalopathy.

Performance of segmentation methods according to lesion characteristics. A, Scatterplot of predicted-versus-true total lesion volume with CNN (green circle) (Spearman correlation ρ = 0.985, best fit line slope β = 0.958), LST (gray square) ρ = 0.862, β = 0.490), and BIANCA (white triangle) (ρ = 0.655, β = 0.277) with the y = x line. Note clustering of CNN points along the y = x line, representing low deviation of CNN-based volume estimates from manual lesion volumes. B, Median Dice scores of cases stratified by total lesion volume. C, False discovery rate stratified by total lesion volume. D, Scatterplot of median CNN Dice score versus median true total lesion volume per diagnostic group. E, Median Dice scores of cases grouped by mean individual lesion volume. F, Histogram of lesion volumes in training and validation datasets. Error bars in all panels represent ±1 standard error of the mean across cases. The asterisk denotes P < .01 for the CNN compared with 1 method, and double asterisks denote P < .01 for CNN compared with both methods using 1-way group ANOVA and paired 2-tailed t tests. The hashtag separately denotes P < .05 for human performance compared with the CNN. ADEM indicates acute disseminated encephalomyelitis; ALD, adrenoleukodystrophy; TLE, toxic leukoencephalopathy; aMS, active MS: tMS, tumefactive MS; PRES, posterior reversible encephalopathy syndrome; iMS, inactive MS; NMO, neuromyelitis optica; Vasc, Vascular disease (ischemia); CNSL, CNS lymphoma; Susac S, Susac syndrome; lg, low-grade; hg, high-grade; PML, progressive multifocal leukoencephalopathy.