Automatically detecting Crohn’s disease and Ulcerative Colitis from endoscopic imaging

Table 2 Metrics values for the best trained nets, ranked by decreasing Matthews Correlation Coefficient (MCC) on the test set; cv, ts: cross-validation and test set; TNR, true negative rate; TPR: true positive rate; NPV, negative predictive value; PPV, positive predictive value; resnetX-Y-Z: meta-model obtained by stacking ensemble of the three weak learners resnetX, resnetY, and resnetZ

Task	Net	MCC_cv	MCC_ts	TNR_ts	TPR_ts	NPV_ts	PPV_ts
N–P	resnet34-50-101	0.976 (0.976,0.976)	0.940	1.000	0.968	0.912	1.000
	resnet50	0.960 (0.954,0.966)	0.937	0.993	0.969	0.915	0.998
	resnet101	0.948 (0.944,0.952)	0.929	0.996	0.964	0.900	0.999
	resnet34	0.954 (0.951,0.956)	0.925	1.000	0.960	0.892	1.000
UC–CD	resnet34-50-152	0.906 (0.905,0.907)	0.688	0.777	0.902	0.875	0.822
	resnet50	0.377 (0.362,0.398)	0.629	0.727	0.890	0.853	0.788
	resnet34	0.428 (0.415,0.454)	0.611	0.744	0.861	0.824	0.793
	resnet152	0.319 (0.319,0.319)	0.602	0.677	0.905	0.862	0.761
UC–N	resnet34-50-101	0.988 (0.988,0.988)	0.931	0.955	0.979	0.983	0.945
	resnet34	0.968 (0.966,0.970)	0.930	0.964	0.968	0.975	0.954
	resnet101	0.964 (0.961,0.970)	0.895	0.967	0.925	0.943	0.956
	resnet50	0.966 (0.962,0.972)	0.855	0.969	0.875	0.909	0.957

ISSN: 1472-6947