Table 7.

Centre-wise performance evaluation of best approach while training using four centres, validating on individual centre, and testing on out-of-sample generalisation task (centre C6).

Method	Val.	JI ↑	DSC ↑	F2 ↑	PPV ↑	Recall ↑	Acc. ↑	dAHD ↓
DeepLabV3+23 (ResNet50)	C1	0.70 ± 0.32	0.76 ± 0.33	0.75 ± 0.33	0.85 ± 0.28	0.79 ± 0.30	0.98	4.77
DeepLabV3+23 (ResNet50)	C2	0.72 ± 0.28	0.79 ± 0.27	0.78 ± 0.27	0.88 ± 0.23	0.80 ± 0.25	0.97	4.70
DeepLabV3+23 (ResNet50)	C3	0.74 ± 0.28	0.80 ± 0.27	0.80 ± 0.26	0.86 ± 0.24	0.83 ± 0.25	0.97	4.77
DeepLabV3+23 (ResNet50)	C4	0.76 ± 0.28	0.82 ± 0.27	0.81 ± 0.28	0.91 ± 0.17	0.81 ± 0.28	0.98	4.77
DeepLabV3+23 (ResNet50)	C5	0.73 ± 0.31	0.78 ± 0.31	0.77 ± 0.32	0.93 ± 0.16	0.78 ± 0.31	0.98	4.88

Val.: Validation centre JI: Jaccard index DSC: Dice coefficient F2: Fbeta-measure, with β = 2 PPV: Positive

predictive valueAcc.: overall accuracy d_AHD: Average Hausdorff distance ↑: best increasing ↓: best decreasing.