Table 2 Cardiac MR segmentation results after continual training measured as an average Dice score (DSC) over LV, RV and MYO segmentation computed on the test set.

Meth.	M	Scanner A	Scanner B	Scanner C	Scanner D	BWT	FWT
DM (Ours)	128	0.802 ± 0.005	0.762 ± 0.002	0.807 ± 0.004	0.840 ± 0.009	0.000 ± 0.002	0.032 ± 0.004
DM-PD (Ours)	128	0.799 ± 0.010	0.763 ± 0.004	0.809 ± 0.005	0.844 ± 0.010	0.003 ± 0.004	0.031 ± 0.005
Random	128	0.786 ± 0.015	0.746 ± 0.008	0.797 ± 0.005	0.847 ± 0.005	−0.011 ± 0.007	0.033 ± 0.004
EWC¹⁸		0.786 ± 0.008	0.738 ± 0.014	0.792 ± 0.007	0.850 ± 0.003	−0.014 ± 0.007	0.032 ± 0.003
Naive		0.781 ± 0.013	0.726 ± 0.026	0.789 ± 0.011	0.848 ± 0.003	−0.018 ± 0.123	0.032 ± 0.002
GEM¹⁹	128	0.798 ± 0.005	0.761 ± 0.008	0.804 ± 0.003	0.846 ± 0.002	−0.005 ± 0.004	0.033 ± 0.003
ER-MIR²⁰	128	0.798 ± 0.005	0.763 ± 0.007	0.808 ± 0.002	0.847 ± 0.001	−0.004 ± 0.003	0.036 ± 0.003
DSM		0.802 ± 0.017	0.748 ± 0.012	0.806 ± 0.014	0.835 ± 0.005	–	–
JModel		0.822 ± 0.010	0.798 ± 0.016	0.823 ± 0.006	0.852 ± 0.007	–	–
Base		0.797	0.763	0.792	0.763	–	–

± 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), random replacement strategy (Random), elastic weight consolidation (EWC) and naive continual learning (Naive). Methods requiring domain membership knowledge are gradient episodic memory (GEM), and experience replay with maximally inferred retrieval (ER-MIR). Domain-specific models (DSM), a joint model (JModel) and using base training only (Base) serve as a reference. For base training only one model was trained to avoid the 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. 1.

Quick links

Search