Skin Lesion Classification Using Additional Patient Information

Table 1

Results of ISIC 2018 challenge winners from the legacy leaderboard (rows 1-3) and our proposed models (rows 4–6). Among the 16,888 images, there are 15,316 images from ISIC19 dataset, 170 images from the MED-NODE dataset, 533 from the seven-point dataset, 120 from the PH2 dataset, and the remaining data are from our own collected data.

Team/authors	Extra images	BMCA (%)	Sensitivity (%)	Specificity (%)	AUC
Nozdryn et al.	37,807	88.5	83.3	98.6	0.983
Gassert et al. [14]	13,475	85.6	80.9	98.4	0.987
MSM-CNN [5]	2,912	86.2	85.6	97.9	0.987
Our single model (FL)	16,888	88.3	76.1	99.3	0.974
Our single model (CE)	16,888	88.7	83.0	98.8	0.981
Our ensemble model	16,888	89.5	89.5	98.1	0.982

FL: focal loss; CE: cross-entropy loss.