Table 1.

Performance comparison of UNETR and SwinUNETR model with different pre-processing method with and without proposed positional embedding (PPE).

Model	Pre-processing Method	Embedding	Mean Dice Score	HD95 (mm)
UNETR	Resample	Regular	0.439	6.53
	Random crop	Regular	0.495	4.74
	Random crop	Regular + PPE	0.560	3.35
SwinUNETR	Resample	Regular	0.537	6.90
	Random crop	Regular	0.557	4.59
	Random crop	Regular + PPE	0.633*	3.04

The HD95 values are computed on true positive components as infarct segmentation has multiple region targets. With proposed embedding and random cropping, the performance improvement is significant.

*Wilcoxon signed-rank test p-value of the best PPE method compared to without PPE methods after Bonferroni correction < 0.0001.