Table 3.

Quantitative results of esophageal GTV segmentation by the pCT+PET model in the unseen internal testing data.

	Institution 1 (Unseen Internal Testing) Using pCT+PET Model
	Unacceptable Number (percentage)	DSC Mean (95% CI)	HD95 (mm) Mean (95% CI)	ASD (mm) Mean (95% CI)
Total patients (n = 104)	4 (4%)	0.83 (0.81, 0.84)	9.5 (8.0, 10.9)	2.2 (1.9, 2.5)
Clinical T stage
cT2 (n = 18)	3 (17%)	0.77 (0.69, 0.85)	11.4 (6.3, 16.6)	2.7 (1.3, 4.2)
cT3 (n = 58)	0 (0%)	0.84 (0.82, 0.85)	9.0 (7.0, 11.0)	2.0 (1.7, 2.4)
cT4 (n = 28)	1 (4%)	0.84 (0.82, 0.86)	9.3 (7.3, 11.4)	2.3 (1.9, 2.6)
Tumor location
Cervical (n = 10)	1 (10%)	0.83 (0.78, 0.89)	9.4 (6.2, 12.7)	2.0 (1.5, 2.5)
Upper third (n = 35)	0 (0%)	0.84 (0.82, 0.86)	8.1 (6.2, 10.0)	1.9 (1.6, 2.2)
Middle third (n = 63)	3 (5%)	0.83 (0.81, 0.84)	9.5 (8.0, 11.1)	2.2 (1.9, 2.5)
Lower third (n = 35)	0 (0%)	0.83 (0.79, 0.86)	10.8 (7.5, 14.0)	2.6 (1.9, 3.3)

GTV, gross tumor volume; CI, confidence interval; DSC, Dice similarity coefficient; HD95, 95% Hausdorff distance; ASD, average surface distance; cT2, clinical T stage 2; cT3, clinical T stage 3; cT4, clinical T stage 4; pCT, treatment planning CT.