CARE-WHS

Whole Heart Segmentation

Motivation

Cardiovascular diseases (CVDs), as the leading cause of death globally, necessitate precise morphological and pathological quantification through segmentation of crucial cardiac structures from medical images. However, whole heart segmentation (WHS) faces challenges including heart shape variability during the cardiac cycle, clinical artifacts like motion and poor contrast-to-noise ratio, as well as domain shifts in multi-center data and the distinct modalities of CT and MRI. The WHS track serves to inspire innovative solutions in the realms of biomedical imaging and computer vision, striving to overcome these challenges and advance automated WHS for enhanced understanding and treatment of CVDs.

Task

The objective of this track is to achieve precise segmentation of seven substructures of the whole heart, with robustness against domain shifts (see Fig. 1).
The specific substructures, each associated with a unique label value, are:

Left Ventricular Blood Cavity (LV) - Label value: 500
Right Ventricular Blood Cavity (RV) - Label value: 600
Left Atrial Blood Cavity (LA) - Label value: 420
Right Atrial Blood Cavity (RA) - Label value: 550
Myocardium of the Left Ventricle (Myo) - Label value: 205
Ascending Aorta (AO) - Label value: 820; defined as the aortic trunk from the aortic valve to the superior level of the atria.
Pulmonary Artery (PA) - Label value: 850; defined as the initial segment from the pulmonary valve to the bifurcation point.

Note on Great Vessels: The great vessels of interest, comprising the ascending aorta and pulmonary artery, are specifically defined due to variations in the fields of view across different scans. This uniform definition is crucial for ensuring consistency across evaluations. During the assessment, segmentation results for these vessels will be truncated to average lengths measured in healthy subjects, although participants are encouraged to extend their segmentation beyond these lengths. Our provided manual segmentations similarly cover more than the defined trunk measurements.

We will rank participant methods based on the settings (Lb1–Lb4) detailed in the following table:

Leaderboard (Lb) for WHS track across modalities and evaluation settings. Lb1–Lb4 represent performance from different test centers and modalities (e.g., Lb1 = CT segmentation peformance on center A and B). In-distribution refers to centers included in the training data, while out-of-distribution refers to unseen centers not used during training.
Modality	In-distribution	Out-of-distribution
CT	Lb1	Lb2
MR	Lb3	Lb4

Data

Training data

Center	Num. patients	Modalities
A	20	CT
B	20	CT
C & D	20	MRI
E	26	MRI

Validation data

Center	Num. patients	Modalities
A	20	CT
B	10	CT
C & D	20	MR

Test data

Center	Num. patients	Modalities
A	20	CT
B	14	CT
C & D	20	MRI
F	16	MRI
G	20	CT

Metrics

The performance of segmentation results will be assessed through:

Dice Similarity Coefficient (Dice)
Hausdorff Distance (HD in mm)
Average Surface Distance (ASD in mm)

Rules

Only automatic methods are acceptable. Participants must utilize algorithms that do not require manual intervention or human-assisted processes for the segmentation task.
Pre-trained models are allowed in this track. The solutions could be developed with pre-trained fundation models, such as SAM, CLIP and MedSAM.

Registration

To access the dataset, please register here.

Leaderboards

WHS CT (ID)

Team	LV_Dice↑	Myo_Dice↑	RV_Dice↑	LA_Dice↑	RA_Dice↑	AO_Dice↑	PA_Dice↑	WHS_Dice↑	WHS_HD(mm)↓	WHS_ASD(mm)↓
wei_yuan	0.9426	0.9300	0.9284	0.9653	0.9240	0.9604	0.8783	0.9410	12.5151	0.7237
dzr	0.9440	0.9294	0.9271	0.9625	0.9208	0.9595	0.8749	0.9400	12.9502	0.7416
huangzh	0.9398	0.9298	0.9239	0.9631	0.9189	0.9601	0.8661	0.9381	12.8066	0.7535
Cardiac Endeavours	0.9421	0.9289	0.9224	0.9646	0.9203	0.9600	0.8630	0.9381	13.1335	0.7525
Nacho	0.9430	0.9272	0.9228	0.9616	0.9184	0.9580	0.8694	0.9378	12.8374	0.7629
spz312	0.9421	0.9270	0.9237	0.9607	0.9167	0.9591	0.8687	0.9374	13.7740	0.7772
Zhiyou Lin	0.9426	0.9242	0.9207	0.9604	0.9193	0.9568	0.8652	0.9361	15.9188	0.7788
Space	0.9344	0.9114	0.9078	0.9617	0.8829	0.9502	0.8490	0.9230	14.8967	0.9549
ImFusion	0.9209	0.8932	0.9041	0.9398	0.9076	0.9206	0.8492	0.9150	17.3939	1.1766
kwh	0.8859	0.7872	0.7052	0.8637	0.6865	0.8949	0.6599	0.8057	36.1450	2.9919

WHS CT (OOD)

Team	LV_Dice↑	Myo_Dice↑	RV_Dice↑	LA_Dice↑	RA_Dice↑	AO_Dice↑	PA_Dice↑	WHS_Dice↑	WHS_HD(mm)↓	WHS_ASD(mm)↓
wei_yuan	0.9754	0.9742	0.9617	0.9839	0.9695	0.9805	0.9561	0.9740	11.6691	0.3616
dzr	0.9772	0.9755	0.9621	0.9830	0.9670	0.9825	0.9527	0.9739	11.9296	0.3668
huangzh	0.9738	0.9719	0.9561	0.9838	0.9653	0.9791	0.9503	0.9713	11.8991	0.3935
Cardiac Endeavours	0.9738	0.9712	0.9553	0.9791	0.9610	0.9757	0.9522	0.9692	16.3845	0.4333
Nacho	0.9749	0.9725	0.9572	0.9802	0.9638	0.9798	0.9527	0.9708	12.1136	0.4074
spz312	0.9763	0.9746	0.9552	0.9841	0.9639	0.9829	0.9482	0.9721	12.3174	0.3887
Zhiyou Lin	0.9721	0.9675	0.9557	0.9806	0.9658	0.9823	0.9526	0.9697	13.8909	0.4089
Space	0.8507	0.9456	0.5362	0.7182	0.8674	0.7533	0.7401	0.8099	48.5065	2.5080
ImFusion	0.9112	0.8980	0.8953	0.9328	0.9235	0.9195	0.8888	0.9135	19.9910	1.2760
kwh	0.9210	0.8349	0.6574	0.9479	0.6643	0.9078	0.7401	0.8257	39.0210	2.7931

The annotation methodology for the gold standards differed across the centers. Datasets from Centers A, B, C, D, and E underwent a fully manual annotation process, while a semi-automatically assisted approach was employed for Center F (OOD).

WHS MR (ID)

Team	LV_Dice↑	Myo_Dice↑	RV_Dice↑	LA_Dice↑	RA_Dice↑	AO_Dice↑	PA_Dice↑	WHS_Dice↑	WHS_HD(mm)↓	WHS_ASD(mm)↓
wei_yuan	0.9354	0.8496	0.9148	0.8959	0.8892	0.8966	0.8148	0.8992	24.7081	1.1629
dzr	0.9329	0.8467	0.9187	0.8937	0.8890	0.8948	0.7973	0.8979	29.6365	1.2133
huangzh	0.9343	0.8503	0.9056	0.8825	0.8725	0.8887	0.8001	0.8913	35.8354	1.6442
Cardiac Endeavours	0.9331	0.8473	0.9115	0.8839	0.8837	0.8957	0.8079	0.8948	27.7563	1.3423
Nacho	0.9323	0.8439	0.9143	0.8862	0.8875	0.8932	0.8068	0.8951	24.0996	1.2457
spz312	0.9164	0.8321	0.9169	0.8927	0.8768	0.8929	0.8099	0.8888	42.4716	2.1076
Zhiyou Lin	0.9357	0.8394	0.9100	0.8817	0.8876	0.8864	0.7815	0.8937	36.7519	1.4858
Space	0.9323	0.8414	0.9008	0.8839	0.8855	0.8891	0.8091	0.8899	29.6204	1.7214
ImFusion	0.4201	0.1910	0.0924	0.3747	0.1258	0.0010	0.0000	0.2183	86.5388	20.5581
kwh	0.9303	0.8404	0.9063	0.8672	0.8612	0.8851	0.7992	0.8852	48.5327	1.9555

WHS MR (OOD)

Team	LV_Dice↑	Myo_Dice↑	RV_Dice↑	LA_Dice↑	RA_Dice↑	AO_Dice↑	PA_Dice↑	WHS_Dice↑	WHS_HD(mm)↓	WHS_ASD(mm)↓
wei_yuan	0.9426	0.8235	0.8735	0.9116	0.8897	0.9106	0.8848	0.8959	20.6044	1.5678
dzr	0.9348	0.8232	0.8742	0.9080	0.8880	0.9112	0.8818	0.8934	21.2980	1.5955
huangzh	0.9462	0.8358	0.8907	0.9056	0.8881	0.9044	0.8766	0.8969	20.7411	1.5648
Cardiac Endeavours	0.9449	0.8329	0.8798	0.9065	0.8903	0.9126	0.8850	0.8965	21.6923	1.5574
Nacho	0.9310	0.8234	0.8772	0.9042	0.8903	0.9077	0.8816	0.8923	21.4862	1.6170
spz312	0.9364	0.8129	0.8644	0.9103	0.8812	0.9058	0.8753	0.8895	22.2377	1.6504
Zhiyou Lin	0.9069	0.7664	0.8718	0.9053	0.8852	0.9018	0.8591	0.8764	28.9588	1.9135
Space	0.9411	0.8355	0.8815	0.9081	0.8862	0.9032	0.8666	0.8947	21.9386	1.5576
ImFusion	0.2172	0.1051	0.0296	0.2463	0.0000	0.0004	0.0000	0.1225	88.9726	26.1159
kwh	0.9017	0.7860	0.8264	0.8997	0.8714	0.8850	0.8484	0.8711	41.6986	2.1071

Citations

Please cite these papers when you use the data for publications:

@article{Zhuang2016MSMMA,
  Author = {Zhuang, Xiahai and Shen, Juan},
  Title = {Multi-scale patch and multi-modality atlases for whole heart
     segmentation of MRI},
  Journal = {Medical Image Analysis},
  Year = {2016},
  Volume = {31},
  Pages = {77-87},
}

@article{Zhuang2019MvMM,
  Author = {Zhuang, Xiahai},
  Title = {Multivariate Mixture Model for Myocardial Segmentation Combining
     Multi-Source Images},
  Journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  Year = {2019},
  Volume = {41},
  Number = {12},
  Pages = {2933-2946},
}

@article{GAO2023BayeSeg,
  Author = {Gao, Shangqi and Zhou, Hangqi and Gao, Yibo and Zhuang, Xiahai},
  Title = {BayeSeg: Bayesian modeling for medical image segmentation with
     interpretable generalizability},
  Journal = {Medical Image Analysis},
  Year = {2023},
  Volume = {89},
  Pages = {102889},
}