Table 3.

Accuracy, ROC AUC, sensitivity and specificity on the validation set for our best model, namely a bidirectional LSTM trained on 120-dimensional feature vectors obtained by applying PCA on the 2048-dimensional features vectors produced by a ResNeXt-101 $32\times 8d$ network. The same metrics are reported for the plain ResNeXt-101 $32\times 8d$ model. Results are report for all intracranial hemorrhage subtypes.

ICH Type	ResNeXt-101 and Bidirectional LSTM
	Accuracy		ROC AUC		Sensitivity		Specificity
	on Slices	on Scans	on Slices	on Scans	on Slices	on Scans	on Slices	on Scans
Any	0.9600	0.9516	0.9843	0.9792	0.8605	0.9436	0.9781	0.9576
Epidural (EPH)	0.9970	0.9892	0.9851	0.9414	0.4430	0.4000	0.9988	0.9973
Intraparenchymal (IPH)	0.9832	0.9368	0.9927	0.9834	0.8142	0.8639	0.9927	0.9620
Intraventricular (IVH)	0.9920	0.9651	0.9970	0.9866	0.9102	0.9127	0.9948	0.9757
Subarachnoid (SAH)	0.9752	0.9220	0.9821	0.9609	0.6624	0.6944	0.9921	0.9767
Subdural (SDH)	0.9627	0.9140	0.9682	0.9451	0.6817	0.7203	0.9837	0.9601
	ResNeXt-101
Any	0.9543	0.9409	0.9752	0.9782	0.7866	0.9436	0.9848	0.9388
Epidural (EPH)	0.9967	0.9879	0.9703	0.9479	0.2405	0.4000	0.9992	0.9959
Intraparenchymal (IPH)	0.9809	0.9449	0.9883	0.9806	0.6865	0.8377	0.9974	0.9819
Intraventricular (IVH)	0.9905	0.9570	0.9953	0.9889	0.8325	0.9127	0.9960	0.9660
Subarachnoid (SAH)	0.9697	0.9126	0.9644	0.9582	0.5024	0.7222	0.9949	0.9583
Subdural (SDH)	0.9604	0.9086	0.9576	0.9408	0.5993	0.8182	0.9873	0.9301