Automated ASPECTS on Noncontrast CT Scans in Patients with Acute Ischemic Stroke Using Machine Learning

H. Kuang; M. Najm; D. Chakraborty; N. Maraj; S.I. Sohn; M. Goyal; M.D. Hill; A.M. Demchuk; B.K. Menon; W. Qiu

doi:10.3174/ajnr.A5889

Article Figures & Data

Figures

Tables

Download figure
Open in new tab
Download powerpoint
Fig 1.
Examples of each ASPECTS region. L indicates lentiform; I, insula; C, caudate; IC, internal capsule; M, MCA.
Download figure
Open in new tab
Download powerpoint
Fig 2.
A flowchart of the training and testing processes used in the study for each ASPECTS region.
Download figure
Open in new tab
Download powerpoint
Fig 3.
A, Boxplot with a scatterplot showing the distribution of the automated CT ASPECTS at each individual ASPECTS on DWI. B, A Bland-Altman plot illustrating agreement between a total automated ASPECTS score and ASPECTS scores on DWI. Random jitter has been added to illustrate the number of measurements at each ASPECTS point. The horizontal black line represents the mean difference in the ASPECTS score between the 2 methods, while the dotted lines represent a 1.96 SD around the difference.
Download figure
Open in new tab
Download powerpoint
Fig 4.
Examples of DWI ASPECTS, the automated CT ASPECTS derived in this study, and expert-read CT ASPECTS. ASPECTS regions with ischemic changes are shown in color.

Tables

Figures

View popup

Table 1:

κ, accuracy, F1 measure, sensitivity, specificity, and AUC on each ASPECTS region

Region	κ (95% CI)	Accuracy (%) (95% CI)	F1 Measure	Sensitivity (%) (95% CI)	Specificity (%) (95% CI)	AUC (95% CI)
M1	0.59 (0.38–0.81)	90 (90/100) (84.1–95.9)	0.64	47.4 (9/19) (24.4–71.1)	100 (81/81) (95.5–100)	0.74 (0.64–0.82)
M2	0.52 (0.35–0.68)	76 (76/100) (67.6–84.4)	0.73	76.2 (32/42) (60.5–87.9)	75.9 (44/58) (62.8–86.1)	0.76 (0.67–0.84)
M3	0.47 (0.21–0.73)	88 (88/100) (81.6–94.4)	0.54	50 (7/14) (23–77)	94.2 (81/86) (87–98.1)	0.72 (0.62–0.81)
M4	0.36 (0.13–0.63)	85 (85/100) (78–92)	0.35	36.4 (4/11) (10.9–69.2)	91.1 (81/89) (83.1–96)	0.64 (0.54–0.73)
M5	0.54 (0.37–0.7)	77 (77/100) (68.8–85.3)	0.74	68.1 (32/47) (52.9–80.9)	84.9 (45/53) (72.4–93.3)	0.77 (0.67–0.84)
M6	0.39 (0.14–0.64)	86 (86/100) (79.2–92.8)	0.46	35.3 (6/17) (14.2–61.7)	96.4 (80/83) (89.8–99.2)	0.66 (0.56–0.75)
Lentiform	0.64 (0.47–0.81)	85 (85/100) (78–92)	0.75	71.0 (22/31) (52–85.8)	91.3 (63/69) (82–96.7)	0.81 (0.72–0.88)
Insula	0.62 (0.46–0.77)	81 (81/100) (73.3–88.7)	0.83	85.5 (47/55) (73.3–93.5)	75.6 (34/45) (60.5–87.1)	0.81 (0.71–0.88)
Caudate	0.63 (0.42–0.84)	90 (90/100) (84.1–95.9)	0.69	57.9 (11/19) (33.5–79.7)	97.5 (79/81) (91.4–99.7)	0.78 (0.68–0.85)
Internal capsule	0.59 (0.35–0.83)	91 (91/100) (85.4–96.6)	0.64	57.1 (8/14) (28.9–82.3)	96.5 (83/86) (90.1–99.3)	0.77 (0.67–0.85)
All regions	0.60 (0.54–0.66)	84.9 (849/1000) (82.7–87.1)	0.70	66.2 (178/269) (60.2–71.8)	91.8 (671/731) (89.6–93.7)	0.79 (0.75–0.83)

View popup

Table 2:

Agreement on ASPECTS interpretation at a regional level and for dichotomized ASPECTS (>4 vs. ≤4) between the automated ASPECTS method and expert-read DWI ASPECTS using different DWI ASPECTS region-involvement thresholds

DWI ASPECTS Region-Involvement Thresholds	κ (95% CI)	Accuracy (%) (95% CI)	F1 Measure	Sensitivity (%) (95% CI)	Specificity (%) (95% CI)	AUC (95% CI)
20%
All regions	0.6 (0.54–0.66)	84.9 (849/1000) (82.7–87.1)	0.70	66.2 (178/269) (60.2–71.8)	91.8 (671/731) (89.6–93.7)	0.79 (0.75–0.83)
>4 and ≤4	0.78 (0.57–0.99)	96 (96/100) (92.2–99.8)	0.98	97.8 (88/90) (92.2–99.7)	80 (8/10) (34.8–93.3)	0.89 (0.81–0.94)
50%
All regions	0.64 (0.57–0.70)	88.8 (888/1000) (86.9–90.8)	0.71	68.2 (133/195) (61.2–74.7)	93.8 (755/805) (91.9–95.4)	0.81 (0.77–0.85)
>4 and ≤4	1 (1–1)	100 (100/100) (100–100)	1	100 (94/94) (96.2–100)	100 (6/6) (54.1–100)	1 (0.96–1)
0%
All regions	0.56 (0.51–0.61)	79.5 (795/1000) (77–82)	0.72	69.7 (264/379) (64.8–74.2)	85.5 (531/621) (82.5–88.2)	0.78 (0.76–0.79)
>4 and ≤4	0.46 (0.26–0.66)	81 (81/100) (73.3–88.7)	0.88	93.2 (68/73) (84.7–97.7)	48.2 (13/27) (28.7–68.1)	0.71 (0.61–0.79)