Feasibility and effectiveness of automatic deep learning network and radiomics models for differentiating tumor stroma ratio in pancreatic ductal adenocarcinoma

Insights into Imaging

Table 4 The performance comparison of different models

Model	Category	Cohort	AUC (95% CI)	ACC	Precision	Recall	F1-Score
Clinical	T stage	Train	0.566 (0.477, 0.654)	0.550	0.694	0.284	0.402
Clinical		Test	0.610 (0.448, 0.772)	0.574	0.769	0.370	0.500
Radiomics	KNeigbors	Train	0.865 (0.832, 0.897)	0.785	0.784	0.785	0.784
		Test	0.717 (0.686, 0.757)	0.702	0.698	0.702	0.699
	SVM	Train	0.925 (0.908, 0.944)	0.847	0.847	0.847	0.847
		Test	0.739 (0.691, 0.791)	0.766	0.761	0.757	0.759
	Logistic regression	Train	0.859 (0.830, 0.886)	0.799	0.798	0.799	0.798
		Test	0.756 (0.719, 0.804)	0.681	0.673	0.670	0.671
	Random forest	Train	0.978 (0.970, 0.987)	0.889	0.891	0.889	0.888
		Test	0.763 (0.725, 0.802)	0.702	0.706	0.676	0.678
Deep learning	ShuffulNet	Train	1.000 (1.000, 1.000)	0.987	0.988	0.988	0.987
		Test	0.846 (0.816, 0.891)	0.830	0.826	0.826	0.826
	Xecption	Train	0.999 (0.999, 1.000)	0.987	0.988	0.988	0.987
		Test	0.924 (0.904, 0.940)	0.851	0.897	0.825	0.860
	MobileNet	Train	0.999 (0.999, 1.000)	0.988	0.988	0.988	0.987
		Test	0.930 (0.911, 0.951)	0.872	0.874	0.882	0.878
	ResNet18	Train	1.000 (1.000, 1.000)	0.998	0.998	0.998	0.998
		Test	0.941 (0.926, 0.962)	0.894	0.904	0.882	0.893