Assistant Diagnosis of Basal Cell Carcinoma and Seborrheic Keratosis in Chinese Population Using Convolutional Neural Network

Table 1

Test results of four CNN models training from scratch and fine-tuning from the pretrained weights. Here, ACC, ACC_Avg, Sen, Spe, and AUC denote accuracy, average accuracy, sensitivity, specificity, and area under the curve, respectively.

Training means	Model	Diagnosis	ACC	ACCAvg	Sen	Spe	AUC
Scratch	InceptionV3	BCC	0.758	0.840	0.852	0.846	0.894
		SK	0.905		0.846	0.852
	InceptionResNetV2	BCC	0.854	0.856	0.854	0.859	0.895
		SK	0.857		0.859	0.854
	DenseNet121	BCC	0.835	0.844	0.846	0.848	0.890
		SK	0.850		0.848	0.846
	ResNet50	BCC	0.786	0.834	0.783	0.883	0.879
		SK	0.872		0.883	0.783
Fine-tuning	InceptionV3	BCC	0.857	0.836	0.852	0.837	0.896
		SK	0.820		0.837	0.852
	InceptionResNetV2	BCC	0.775	0.855	0.791	0.915	0.919
		SK	0.918		0.915	0.791
	DenseNet121	BCC	0.871	0.846	0.885	0.816	0.913
		SK	0.826		0.816	0.885
	ResNet50	BCC	0.758	0.845	0.808	0.894	0.905
		SK	0.913		0.894	0.808