Table 1.

Review of the Last Segmentation Methodologies and Winners of the ACDC 2017 Challenge.

Reference	Dataset						Model						CV Method	Reported Result (%)^* (LVC, RVC, LVM, Mean)
Reference	Dataset Name	Sample Size (trains/tests)	Class Size	Pre processing	Data Augment.	Crop-Resize	Localizer	Base Model	Attention mechanism	Additional data	Loss Function (s)	Ex. Classifier	CV Method	Reported Result (%)^* (LVC, RVC, LVM, Mean)

Isensee et al. 2017 (Isensee et al., 2017)	ACDC-17	100/50	4	-	✓	-	-	3D U-Net + 2D U-Net	-	-	C+D	-	5-fold	94.5, 90.8, 90.5, 91.9
Khened et al. 2017 (Khened et al., 2017)	ACDC-17	100/50	4	✓	✓	-	✓	Densely FCN	-	✓	D	RF	H.O (90 %)	94.1, 90.7, 89.4, 91.3
Zotti et al. 2018 (Zotti et al., 2018)	ACDC-17	100/50	4	-	-	-	-	Grid net	-	-	C	-	H.O (75 %)	93.8, 91.0, 89.4, 91.4
Painchaud et al. 2020 (Painchaud et al., 2020)	ACDC-17	100/50	4	-	✓	-	-	VAE	-	✓	KL loss	-	10-fold	93.6, 90.9, 88.9, 91.1
Shi et al. 2021 (Shi et al., 2021)	ACDC-17 LV-09 RV-12	100 /50 130/20 80/10	4 4 4	✓	-	-	✓	MIFNet	PAM	-	D	-	4-fold	Mean: 97.2 Mean: 96.1 Mean: 89.6
Zhou et al. 2021 (Zhou et al., 2021)	ACDC-17	100/50	4	✓	✓	-	-	nnFormer	Multi-head self-attention	-	MSE	-	H.O (70 %)	Mean: 92.1
Chen et al. 2022 (Chen et al., 2022)	ACDC-17 2018 ASC	100/50 50	4 5	✓	-	-	-	Multiresolution Aggregation Transformer U-Net	Coordinate attention (CA)	-	C+D	-	H.O (70 %)	Mean: 92.0 Mean: 92.1
Galazis et al. 2021 (Galazis et al., 2021)	ACDC-17	100/50	4	✓	✓	-	✓	Tempera: Spatial Transformer Feature Pyramid	PAM	✓	D	-	H.O (90 %)	Mean: 92.1
Fu et al. 2022 (Fu et al., 2022)	ACDC-17 Synapse	100/50 50	4 8	✓	-	✓	-	TF-U-Net	-	-	C	-	H.O (80 %)	Mean: 91.7 Mean: 85.5
D. Li et al. 2022 (Li et al., 2022a)	ACDC-17	100/50	4	✓	✓	✓	-	U-Net	PAM CAM	-	C+D	-	H.O (80 %)	RV: 89.1
Y. Li et al. 2022 (Li et al., 2022b)	ACDC-17	100/50	4	✓	-	-	-	ViT Decoder	Window Attention Unsampled	-	MSE	-	H.O (80 %)	Mean: 92.1
Galea et al. 2021 (Galea et al., 2021)	2018 ASC	50	5	✓	✓	-	-	U-Net + DeepLabV3+	-	-	MSE	-	H.O (80 %)	Mean: 92.1
Garcia-Cabrera et al. 2023 (Garcia-Cabrera et al., 2023)	ACDC-17	100/50	4	-	✓	-	-	U-Net (ImageNet)	-	-	MSE	-	H.O (80 %)	Mean:96.6
Płotka et al. 2023 (Grzeszczyk et al., 2023)	ACDC-17	100/50	4	-	-	-	-	Swin Transformer/U-Net	Shifted-window multi-head	-	C+D	-	5-fold	Mean:87.1
Liu et al. 2023 (Liu et al., 2023)	ACDC-17	100/50	4	-	-	-	-	Successive subspace learning	-	-	MSE	-	5-fold	Mean:95.0
J.M.Harana et al. 2021 (Mariscal-Harana et al., 2021)	ACDC-17 M&Ms-2 Local (NHS)	100/50 360 (321 MRI) 4228	4	✓	-	-	-	2D U-Net	-	-	C+D	-	5-fold	87.8, 81.7, 79.1, 82.9 83.8, 80.7, 82.2, 82.2 90.1, 82.1, 78.9, 83.7
A.K. Yasmina et al. 2023 (Al Khalil et al., 2023)	M&Ms-2	360 (200 annotated)	4	-	✓	-	-	Conditional GAN	Post processing	-	D+H	-	5-fold	95.9, 90.7, 93.8, 93.4
C. Galazis et al. 2023 (Galazis et al., 2021)	M&Ms-2	360 (200 annotated)	4	✓	✓	✓	✓	Tempera	Post processing	-	F+D	-	H.O (93 %)	Mean: 83.6
Gao et al. 2022 (Gao and Zhuang, 2022)	M&MS-2	360 (200 annotated)	4	✓	✓	-	✓	U-Net+ ViT	Self-attention	-	C+D	-	H.O (80 %)	Mean: 88.1
Habijan et al. 2021 (Habijan et al., 2021)	MM-WHS	100/40	7	✓	✓	-	-	VAE and FM-Pre- ResNet	-	-	D, L2, KL	-	10-fold	Mean: 89.5
Yang et al. 2022 (Yang et al., 2022a)	Synapse	50	8	✓	-	-	-	Encoder-Decoder	Fuzzy	-	-	-	H.O (80 %)	Mean: 85.5
Pereira et al. 2020 (Pereira et al., 2020)	MICCAI 2019	1120 +56	4	-	-	-	✓	U-Net	-	-	MAE	-	5-fold	Mean: 94.2

ACDC-17: Automated Cardiac Diagnosis Challenge dataset, ACC: Accuracy, C: Cross-Entropy Loss, CAM: Channel Attention Module, D: Dice Loss, H: Hausdorff Distance, H.O: Hold Out, KL: Kullback-Leibler Divergence, L2: L2 norm, LV: Left Ventricle, LV-09: Sunnybrook Cardiac dataset, Myo: Myocardium, PAM: Position Attention Module, RF: Random Forest, ROI: Region of Interest, RV: Right Ventricle, RV-12: MICCAI 2012 Right Ventricle Segmentation Challenge dataset, VAE: Variational Autoencoder.

The results are reported based on CMRI analyses like LV, RV, Myo, and Overall mean, respectively. Here, class size indicates the segmented heart regions count plus the background.