Table 2.

CNN Models’ performance in internal test and external test

		Source	DSCs	Jaccard	TPR	FPR	95HD	Volume ICC
Classic U-Net	Internal test	Center A	0.767 (0.701~0.833)	0.647 (0.573~0.721)	0.802 (0.713~0.890)	0.001 (0.0004~0.0008)	5.726 (3.860~7.591)	0.977 (0.951~0.990)
	External test	Center B	0.754 (0.679~0.829)	0.646 (0.571~0.721)	0.736 (0.651~0.822)	0.000 (0.0004~0.0006)	6.676 (4.695~8.657)	0.954 (0.829~0.982)
Modified U-net	Internal test	Center A	0.840 (0.803~0.877)	0.734 (0.685~0.782)	0.820 (0.770~0.870)	0.000 (0.0002~0.0004)	3.669 (2.534~4.803)	0.987 (0.966~0.994)
	External test	Center B	0.816 (0.773~0.858)	0.704 (0.650~0.761)	0.765 (0.703~0.828)	0.000 (0.0002~0.0004)	4.201 (3.294~5.109)	0.978 (0.957~0.989)

DSCs dice similarity coefficients, 95HD Hausdorff surface distance of 95% percentile, TPR true positive rate, FPR false positive rate, ICC interclass correlation coefficients