Ensemble of Convolutional Neural Networks Improves Automated Segmentation of Acute Ischemic Lesions Using Multiparametric Diffusion-Weighted MRI

Article Figures & Data

Download figure
Open in new tab
Download powerpoint
Fig 1.
Median Dice (80.2% [IQR, 56.6%–88.9%]), precision (82.9% [IQR, 59.7%–92.2%]), and sensitivity (86.2% [IQR, 71.1%–92.3%]) scores of the DWI+ADC+LOWB ensemble on the Evaluation Cohort. The white bar within the violin plot shows the IQR, mean is a diamond, and median is an X.
Download figure
Open in new tab
Download powerpoint
Fig 2.
Sample segmentation results of the ensemble of DWI+ADC+LOWB (blue regions) on sample subjects along with manual outlines (red outlines). A, A small lesion example from a 70-year-old man with an admission NIHSS score of 1, imaged approximately 9 hours from LKW: MLV = 0.96 cm³, ALV = 1.07 cm³, Dice = 89.4%. B, Medium lesion sample from a 38-year-old woman with an admission NIHSS score of 4, imaged approximately 10 hours from LKW: MLV = 54.3 cm³, ALV = 57.9 cm³, Dice = 95.7%. C, A large lesion example from a 62-year-old man with an undocumented admission NIHSS score, imaged approximately 10 hours from LKW: MLV = 229.0 cm³, ALV = 208.7 cm³, Dice = 94.0%.

Table 1:

Demographics for training and Evaluation Cohorts^a

Characteristic	Training (n = 116)	Evaluation (n = 151)	P Value
Age (yr)	67.9 ± 17.2	65.2 ± 15.5	0.11
Male sex	57 (49.1%)	104 (68.9%)	.002
NIHSS score	7 (3–15.75)^b	6 (3–13)^c	.53
Time to MRI (h)	5.0 (2.9–6.8)	6.2 (3.8–8.3)	.002
Manual lesion volumes (cm³)	9.0 (1.5–28.4)	10.6 (2.0–32.4)	.60

↵a Differences as a factor of the Training Cohort are shown. Data are shown as median (IQR), mean ± SD, or No. (%).
↵b n = 112.
↵c n = 115.

Table 2:

Comparison of performance metrics of segmentations for different CNN models^a

Model	Dice	Precision	Sensitivity
LOWB	6.5 (0.3–20.9)	5.7 (0.3–32.7)	8.5 (0.3–28.5)
ADC^b	56.4 (27.1–75.4)	59.4 (22.3–78.4)	58.2 (32.7–78.9)
DWI	72.3 (46.2–82.5)	73.0 (38.3–88.1)	84.0 (62.4–90.8)
ADC+LOWB	76.5 (51.9–86.1)	78.1 (47.2–88.8)	79.2 (66.6–89.7)
DWI+LOWB	76.7 (58.4–85.4)	79.4 (52.0–89.8)	83.0 (64.8–90.6)
DWI+ADC	79.0 (57.1–86.4)	79.0 (62.1–90.5)	82.6 (68.4–91.4)
DWI+ADC+LOWB	78.9 (56.2–86.2)	77.4 (55.0–89.8)	83.4 (71.3–91.8)
E2 (DWI+ADC)	82.0 (62.9–88.1)	82.0 (65.1–92.6)^b	84.1 (71.0–92.6)
E3 (DWI+ADC+LOWB)	82.2 (64.9–88.9)	83.2 (67.7–93.3)	83.9 (71.9–92.4)

↵a All metrics are denoted in percentages as median (IQR). Of the nonensemble models, significant differences in Dice, precision, and sensitivity were found (P < .001). The ensemble models, E2 and E3, were superior to all other models (P < .001).
↵b Excludes 1 subject with an automatically segmented lesion volume of zero because precision is undefined in this circumstance.

Table 3:

Dependency of automated segmentation performance on MLV^a

Group	Thresholds	Dice	Precision^b	Sensitivity	Correlation
I-A	MLV < 1 cm³ (n = 22)	31.0 (0–50.0)	29.6 (3.4–54.9)	57.5 (0–90.0)	ρ = 0.09, P = .68
I-B	MLV ≥ 1 cm³ (n = 129)	83.5^c (71.2–89.3)	84.9^c (70.3–92.9)	87.6^d (75.8–92.9)	ρ = 0.90, P < .001
II-A	MLV < 21 cm³ (n = 100)	71.2 (45.8–84.8)	71.6 (38.7–84.9)	81.3 (59.8–92.5)	ρ = 0.79, P < .001
II-B	MLV ≥ 21 cm³ (n = 51)	89.4^c (85.4–92.5)	92.3^c (85.6–96.1)	89.3^e (83.0–92.2)	ρ = 0.97, P < .001
III-A	MLV < 31 cm³ (n = 113)	73.6 (48.0–85.8)	77.2 (46.3–85.7)	82.5 (62.8–92.1)	ρ = 0.83, P < .001
III-B	MLV ≥ 31 cm³ (n = 38)	90.6^c (87.3–93.2)	94.7^c (88.4–96.8)	89.4^e (82.8–93.6)	ρ = 0.96, P < .001
IV-A	MLV < 51 cm³ (n = 124)	75.0 (48.9–86.8)	78.1 (49.2–86.5)	83.3 (65.2–92.5)	ρ = 0.87, P < .001
IV-B	MLV ≥ 51 cm³ (n = 27)	91.5^c (89.1–93.6)	95.9^c (92.2–97.5)	89.2 (83.5–92.2)	ρ = 0.92, P < .001
V-A	MLV < 70 cm³ (n = 131)	77.2 (51.5–87.0)	79.9 (54.2–87.0)	84.0 (67.8–92.6)	ρ = 0.88, P < .001
V-B	MLV ≥ 70 cm³ (n = 20)	91.8^c (89.4–93.9)	96.0 (93.0–96.9)	89.6 (85.0–92.0)	ρ = 0.83, P < .001

↵a Performance metrics are in median (IQR) and percentages. Results of E3 applied to the Evaluation Cohort are shown as a function of different volume thresholds.
↵b Excludes 2 subjects in group A with automatically segmented lesion volumes of zero because precision is undefined in this circumstance.
↵c P < .001.
↵d P < .01.
↵e P < .05 group A versus group B, where Group A is the group meeting the threshold criteria and Group B is the group not meeting the threshold criteria.