Table 3.

CT Lung nodule detection results after continual training measured in average precision (AP) computed on the test set.

Meth.	M	Scanner E	Scanner F	Scanner G	Scanner H	BWT	FWT
DM (Ours)	128	0.722 ± 0.020	0.526 ± 0.021	0.592 ± 0.041	0.330 ± 0.015	0.030 ± 0.018	0.063 ± 0.016
DM-PD (Ours)	128	0.750 ± 0.006	0.565 ± 0.067	0.624 ± 0.024	0.355 ± 0.038	0.028 ± 0.019	0.066 ± 0.030
Random	128	0.752 ± 0.019	0.514 ± 0.021	0.600 ± 0.021	0.394 ± 0.013	0.007 ± 0.016	0.084 ± 0.026
Naive		0.682 ± 0.014	0.506 ± 0.017	0.561 ± 0.020	0.369 ± 0.008	0.000 ± 0.008	0.091 ± 0.027
GEM19	128	0.754 ± 0.012	0.568 ± 0.022	0.622 ± 0.038	0.366 ± 0.024	0.034 ± 0.016	0.067 ± 0.018
ER-MIR20	128	0.754 ± 0.012	0.588 ± 0.038	0.611 ± 0.039	0.363 ± 0.027	0.031 ± 0.016	0.075 ± 0.016
DSM		0.653 ± 0.047	0.441 ± 0.074	0.643 ± 0.067	0.454 ± 0.096	–	–
JModel		0.716 ± 0.063	0.522 ± 0.114	0.711 ± 0.058	0.419 ± 0.087	–	–
Base		0.645	0.372	0.509	0.136	–	–

± indicates the interval over n = 5 independent runs with different seeds. Dynamic memory (DM) is compared to DM with a pseudo-domain module (DM-PD), naive continual learning, random replacement strategy (Random), domain-specific models (DSM), a joint model (JModel) and using base training only (Base). In addition, GEM and ER-MIR are shown for reference, noting that they require information about domain membership. For base training only one model was trained to avoid influence of base training results on subsequent continual training, therefore no standard deviations are indicated. For a visual presentation of the results, see Supplementary Fig. 2.