Table 2 Classification performance metrics of different models trained from the beginning in the test set

Model	Class	AUC	Accuracy	Sensitivity	Specificity	PPV	NPV
Prognose-CNN (without Attention)	Response	0.99 ± 0.0012	0.8476 ± 0.0082	0.9189 ± 0.0071	0.9618 ± 0.0012	0.9315 ± 0.0001	0.9124 ± 0.0022
	Stable	0.92 ± 0.0014		0.7903 ± 0.0113	0.8741 ± 0.0007	0.7313 ± 0.0002	0.9058 ± 0.0043
	Non-Response	0.97 ± 0.0015		0.8260 ± 0.0009	0.9412 ± 0.0010	0.8769 ± 0.0034	0.9143 ± 0.0007
Prognose-CNNattention	Response	0.99 ± 0.0021	0.8692 ± 0.0071	0.9595 ± 0.0011	0.9618 ± 0.0014	0.9342 ± 0.0026	0.9767 ± 0.0113
	Stable	0.93 ± 0.0012		0.7749 ± 0.0026	0.9090 ± 0.0023	0.7869 ± 0.0113	0.9028 ± 0.0012
	Non-Response	0.97 ± 0.0023		0.8550 ± 0.0022	0.9338 ± 0.0026	0.8676 ± 0.0011	0.9270 ± 0.0009
Auto-Prognose-CNNattention (SegforClass)	Response	0.98 ± 0.0014	0.8309 ± 0.0113	0.8920 ± 0.0026	0.9542 ± 0.0025	0.9167 ± 0.0012	0.9398 ± 0.0002
	Stable	0.89 ± 0.0016		0.8065 ± 0.0007	0.8392 ± 0.0009	0.6849 ± 0.0071	0.9091 ± 0.0011
	Non-Response	0.96 ± 0.0017		0.7826 ± 0.0082	0.9559 ± 0.0020	0.9000 ± 0.0043	0.8966 ± 0.0022

AUC sensitivity, specificity, PPV and NPV metrics have different values per class following the one-vs-rest scheme. Performance improvement is observed for the architecture that includes attention mechanism. Auto-Prognose-CNNattention (SegforClass) framework consists of the tumor segmentation model followed by the prognostic prediction model.

Quick links

Search