Table 1.

Overall performance of CNN features at the default probability threshold 0.5. ft indicates fine-tuned model. The ft-CNN model utilizes the fine-tuned CNN architecture as an end-to-end classifier; while all other models use either handcrafted or CNN features to train classifiers. This table lists the means ± standard deviations of our ten-fold ten-cross validation results.

Model	AUC(%)	accu(%)	sensi(%)	speci(%)
SVM.PLBP-PLAB-PHOG	80.71±6.15	77.17±6.62	78.55±6.17	75.80±8.39
SVM.CNN-fc6	69.81±5.02	66.01±3.10	65.07±5.52	66.96±6.79
SVM.CNN-fc7	75.05±5.50	69.13±5.16	69.57±8.08	68.70±7.29
SVM.ft-CNN-fc6	79.78±4.60	74.20±4.65	75.36±7.48	73.04±6.55
SVM.ft-CNN-fc7	80.01±4.99	74.64±5.71	76.52±9.11	72.75±6.09
ADA.ft-CNN-fc7	80.30±4.07	77.39±3.89	80.87±6.69	73.91±9.23
ft-CNN	82.31±4.63	78.41±5.01	80.87±7.43	75.94±7.46