Table 2.

The aggregated DSC (%) values of ablation study of our method, based on 5-fold cross-validation for Task-2. “Base” refers to fully supervised learning using mid-RT as inputs. “Base+pre-RT” signifies that the inputs have been expanded to include both mid-RT and registered pre-RT along with their labels. “DFUNet” indicates the substitution of the model with the DFUNet architecture compared to “Base+pre-RT”. “Pre-train+MixUp” means the experiments were conducted on the basic segmentation network depicted in Fig.1 (a) with pre-trained weights and MixUp augmentation.

	Base		Base+pre-RT		DFUNet		Pre-train+MixUp
	GTVp	GTVn	GTVp	GTVn	GTVp	GTVn	GTVp	GTVn
Fold1	40.46	74.10	59.80	86.62	64.48	86.82	62.08	87.63
Fold2	27.34	66.65	59.01	85.00	56.93	85.36	56.91	84.77
Fold3	38.21	75.47	65.46	86.96	58.74	86.65	64.98	87.45
Fold4	36.15	70.32	63.65	88.09	65.85	87.14	64.89	88.32
Fold5	45.85	79.72	58.66	86.59	55.87	86.92	58.24	87.55
Average	37.60	73.25	61.32	86.65	60.37	86.58	61.42	87.14