A Deep Learning Segmentation Approach in Free-Breathing Real-Time Cardiac Magnetic Resonance Imaging

Table 1

Comparing the segmentation performance of the proposed network and U-Net on validation and testing set (IoU: intersection over union, CE: cross-entropy, IGD: improved generalized Dice, DA: data augmentation, and italic formatting indicating the best Dice and IoU score for each class).

Loss function	LV Dice		LV Myocardium Dice		RV Dice
	valid	test	valid	test	valid	test
U-Net +CE	0.763	0.815	0.632	0.690	0.675	0.680
U-Net +CE+DA	0.867	0.880	0.757	0.782	0.758	0.746
U-Net +IGD	0.784	0.835	0.676	0.722	0.741	0.684
U-Net +IGD+DA	0.870	0.876	0.752	0.771	0.768	0.788
Proposed+CE	0.765	0.815	0.646	0.701	0.722	0.704
Proposed+CE+DA	0.823	0.875	0.704	0.770	0.767	0.748
Proposed+IGD	0.804	0.817	0.668	0.711	0.730	0.670
Proposed+IGD+DA	0.878	0.919	0.768	0.806	0.795	0.818
	LV IoU		LV Myocardium IoU		RV IoU
	valid	test	valid	test	valid	test
U-Net +CE	0.697	0.758	0.517	0.585	0.606	0.615
U-Net +CE+DA	0.798	0.827	0.640	0.679	0.702	0.694
U-Net +IGD	0.712	0.772	0.558	0.615	0.677	0.620
U-Net +IGD+DA	0.791	0.816	0.618	0.661	0.715	0.734
Proposed+CE	0.709	0.760	0.531	0.592	0.651	0.639
Proposed+CE+DA	0.753	0.822	0.590	0.669	0.710	0.697
Proposed+IGD	0.733	0.757	0.550	0.602	0.659	0.605
Proposed+IGD+DA	0.802	0.860	0.642	0.699	0.741	0.761