Table 2.

Reproducibility results on BraTS 2015 presented in the original paper for the 3D dual-path CNN⁹ and for the 2D single-path CNN¹⁰ (original) and for our independent reproducibility analysis (this work)

	Dice similarity coefficient			Positive predictive value			Sensitivity
	Whole tumour	Tumour core	CE tumour	Whole tumour	Tumour core	CE tumour	Whole tumour	Tumour core	CE tumour
3D dual-path CNN
Original	0.85	0.67	0.63	0.85	0.85	0.63	0.88	0.61	0.66
This work	0.85	0.68	0.64	0.85	0.83	0.62	0.88	0.64	0.70
2D single-path CNN
Original	0.78	0.65	0.75	–	–	–	–	–	–
This work (HGG)	0.36	0.25	0.17	0.36	0.21	0.29	0.54	0.58	0.17
This work (LGG)	0.25	0.14	0.13	0.40	0.51	0.37	0.25	0.10	0.10

Our analysis was carried out for HGG and LGG model parameters of the 2D single-path CNN. The results were congruent with the original analysis for the 3D dual-path CNN but they show an unsuccessful attempt to reproduce the 2D single-path CNN validation. The higher score in each column is emphasised in bold. Measures of dispersion or significance of differences were not available for the original method evaluation.

BraTS, Brain Tumour Segmentation Challenge; CE, contrast-enhanced; CNN, convolutional neural network; 2D, two dimensions; 3D, three dimensions; HGG, high-grade glioma; LGG, low-grade glioma.