Gastric Pathology Image Classification Using Stepwise Fine-Tuning for Deep Neural Networks

Table 3

Performances of the proposed two-stage fine-tuning using cell-wise (nuclei) data.

Data size	Scheme	CNN architecture
		VGG-16				AlexNet				GoogLeNet (Inception V3)
		AUC	ACC	Precision	Recall	AUC	ACC	Precision	Recall	AUC	ACC	Precision	Recall
Small	One stage	0.879	0.793	0.863	0.695	0.828	0.723	0.74	0.72	0.838	0.753	0.75	0.75
	Two stage (nuclei)	0.902	0.815	0.873	0.730	0.852	0.777	0.78	0.78	0.872	0.784	0.78	0.78
Large	One stage	0.936	0.836	0.957	0.703	0.867	0.794	0.80	0.79	0.881	0.779	0.79	0.78
	Two stage (nuclei)	0.965	0.89	0.881	0.901	0.915	0.834	0.84	0.83	0.933	0.862	0.86	0.86