. 2022 Apr 28;35:103027. doi: 10.1016/j.nicl.2022.103027

Table 3.

Model performance.

Model	Sensitivity¹	FP_TBI²	FP^healthy2	FP_all²/ CMB^RS	MAE^TBI	nMAE^TBI	MAE_all	nMAE_all
Baseline	$80.3 %$ $(13.7 %)$	$32.1$ $(10.2)$	$29.3$ $(17.1)$	$3.31$	$52.1$ $(17.9)$	$2.89$	$40.7$ $(20.5)$	$4.52$
Patch-CNN	$87.7 %$ $(9.9 %)$	$20.6$ $(9.8)$	$6.9$ $(3.8)$	$1.51$	$61.2$ $(30.8)$	$3.4$	$34.1$ $(34.9)$	$3.78$
Segmentation-CNN	$91.0 %$ $(8.1 %)$	$19.2$ $(6.1)$	$5.5$ $(3.1)$	$1.27$	$35.1$ $(10.7)$	$1.95$	$20.3$ $(16.8)$	$2.26$
U-Net	$92.2 %$ $(8.3 %)$	$17.1$ $(4.7)$	$3.4$ $(2.1)$	$1.08$	$31.2$ $(8.6)$	$1.73$	$17.3$ $(15.2)$	$1.92$

All values are presented as mean (standard deviation) on the test set (not bootstrapped).

¹ Sensitivity is calculated an operating point of $10$ FPs averaged per TBI patient.

² FP count is calculated at an operating point of $90 %$ average sensitivity.

FP: False positive, TBI: Traumatic Brain Injury, CMB: Cerebral Microbleed, RS: Reference Standard (manually corrected majority vote), MAE: Mean Absolute Error, nMAE: normalized Mean Absolute Error.