Table 5.

Quantitative results of nuclear segmentation

Method	Task1 (liver)		Task2 (breast)		Task3 (kidney)		Task4 (prostate)
	Dice ↑	AJI ↑	Dice ↑	AJI ↑	Dice ↑	AJI ↑	Dice ↑	AJI ↑
JCL	0.6676	0.3420	0.7114	0.4457	0.7350	0.4814	0.7627	0.5184
Fine-Tuning	0.6676	0.3420	0.6950	0.4405	0.7142	0.4195	0.6902	0.4273
TDGAN	0.6676	0.3420	0.6961	0.4323	0.7164	0.4512	0.7481	0.4931
CL-DSL	0.6676	0.3420	0.7346	0.4638	0.7428	0.4605	0.7633	0.4828

All comparing methods learn GAN models in the continual learning setting and the segmentation models are trained from the corresponding synthetic data. The reported results include the mean of the Dice score and Aggregated Jaccard Index (AJI). Note that the results in the Task1 columns are the same among all the methods because the learning processes of Task1 are the same for different approaches. CL-DSL outperforms the other methods and obtains close performance to the centralized-learning method JCL.