Article Figures & Data
Figures
- FIG 1.
Violin plots demonstrating the distribution of Dice scores for nnU-Net and DeepMedic segmentation compared with ground truth for both internal and external test sets. A, WT. B, ET. C, NET. D, CC. E, ED. F, All NETs (NET + CC + ED).
- FIG 2.
Scatterplots of the correlations between ground truth and automated tumor subregion volume percentages from nnU-Net and DeepMedic for internal and external test sets. A and C, Proportion of tumor that is labeled enhancing, B and D, Proportion of tumor that is labeled NET.
- FIG 3.
Bland-Altman analysis plots demonstrating the agreement between ground truth and automated tumor subregion volumes/percentages from nnU-Net and DeepMedic for internal and external test sets. A and C, Proportion of tumor that is labeled enhancing. B and D, Proportion of tumor that is labeled NET.
- FIG 4.
Sample comparison images of ground truth and predicted segmentation from nnU-Net and DeepMedic in a predominantly cystic supratentorial mass with a solid enhancing nodule. While nnU-Net has near-perfect agreement with the ground truth, DeepMedic failed to delineate most of the large cyst, enhancing nodule, and nonenhancing rim. T1w indicates T1-weighted; T1-GD, T1-gadolinium.
Tables
- Table 1:
Characteristics of patients, tumor histology, and MR imaging scanners in internal and external patient cohorts
Internal Cohort External Cohort Total patients 293 46 Median age at imaging (range) (yr) 7.84 (0.24–21.71) 9.42 (0.55–20.73) Sex Male 157 (53.5%) 22 (47.8%) Female 134 (45.7%) 24 (52.1%) NA 2 (0.6%) Histology Low-grade glioma/astrocytoma 152 (51.9%) 21 (45.7%) High-grade glioma/astrocytoma 23 (7.8%) 3 (6.5%) Ependymoma 7 (2.4%) 0 (0%) Medulloblastoma 85 (29%) 17 (36.9%) Brainstem glioma 15 (5.1%) 2 (4.3%) Germinoma 2 (0.7%) 0 (0%) Other 9 (3.1%) 3 (6.5%) Percentage of tumors containing each subregion ET 85.0% 91.3% NET 95.0% 93.5% CC 68.3% 84.8% ED 43.3% 58.7% Note:—NA indicates not available.
- Table 2:
Results comparing the performance metrics of nnU-Net versus DeepMedic architectures for WT and tumor component segmentations compared with the manual ground truth in terms of Dice score metric
Region Internal Test Subjects External Test Subjects Dice Score: Mean (median) Dice Score: Mean (median) nnU-Net DeepMedic nnU-Net DeepMedic WT 0.9 (SD,0.07) (0.94) 0.82 (SD, 0.16) (0.88) 0.88 (SD, 0.07) (0.9) 0.78 (SD, 0.18) (0.86) ET 0.77 (SD, 0.29) (0.86) 0.66 (SD, 0.32) (0.75) 0.75 (SD, 0.26) (0.85) 0.65 (SD, 0.32) (0.8) NET 0.66 (SD, 0.32) (0.80) 0.48 (SD, 0.27) (0.49) 0.53 (SD, 0.32) (0.64) 0.4 (SD, 0.27) (0.4) CC 0.71 (SD, 0.33) (0.83) 0.48 (SD, 0.36) (0.55) 0.55 (SD, 0.33) (0.67) 0.37 (SD, 0.33) (0.35) ED 0.71 (SD, 0.40) (1) 0.19 (SD, 0.33) (0) 0.4 (SD, 0.43) (0.19) 0.21 (SD, 0.32) (0) NET+CC+ED 0.82 (SD, 0.14) (0.86) 0.67 (SD, 0.2) (0.73) 0.74 (SD, 0.2) (0.8) 0.59 (SD, 0.23) (0.65) TC 0.91 (SD, 0.07) (0.94) 0.78 (SD, 0.18) (0.83) 0.87 (SD, 0.07) (0.88) 0.76 (SD, 0.18) (0.79) - Table 3:
Results on the performance of our trained autosegmentation model using nnU-Net architecture on multi-institutional BraTS-PEDs 2023 data (n = 92)a
Region/Metric Dice Score Mean (Median) Sensitivity Mean (Median) 95% Hausdorff Distance Mean (Median) WT 0.87 (SD, 0.13) (0.91) 0.82 (SD, 0.16) (0.88) 6.47 (SD, 15.99) (3) ET 0.48 (SD, 0.38) (0.58) 0.55 (SD, 0.35) (0.71) 11.26 (SD, 20.06) (3.40) TC 0.83 (SD, 0.18) (0.89) 0.82 (SD, 0.17) (0.87) 7.65 (SD, 16.27) (3.67) Note:—The metrics are reported for whole tumor, enhancing tumor, and tumor core (ET + CC + NET) in agreement with BraTS-PEDs evaluation method.
