Left ventricle (LV) function parameters such as stroke volume and ejection fraction (EF) are vital prognostic indicators in the management of heart diseases. Currently these parameters are derived from multiple short-axial cine images acquired under breath hold, which requires extensive scan planning and patient cooperation. The post-processing workflow is also labor-intensive as a human operator is required to manually trace the endocardium border in every slice. Also, the calculated LV volume may be inaccurate due to low slice resolution and slice mismatch from inconsistent breath hold positions. In this work, we propose to combine a recently developed free-breathing 4D MRI technique and atlas-based image segmentation to calculate the LV function parameters fully automatically.MR data were collected using a self-gated, contrast-enhanced, spoiled gradient echo pulse sequence with 3D radial k-space trajectory on a 3T clinical scanner (MAGNETOM Verio, Siemens Healthcare, Germany). The scan parameters are as follows: TR/TE=6.0/3.7 ms, flip angle = 15°, FOV = (320 mm)3, matrix size = 3203, number of lines = 99,994, scan time = 10 minutes, contrast enhancement with 0.20 mmol/kg Gd-BOPTA (MultiHance, Bracco) injected at 0.3 mL/s before acquisition. From each dataset, a 16-phase whole-heart cine series was reconstructed using the methods proposed in [1]. Healthy volunteers (N=5) were scanned with informed consent and IRB approval. The LV and myocardium from all subjects were labeled manually as both the reference standard and the basis for our atlas. For automated processing, all images first went through N4 bias correction to remove coil sensitivity bias. Then, preprocessing was performed to generate additional inputs for image registration, including SUSAN denoising, and Laplacian edge enhancement [2]. Next, an unbiased template was created for each subject through motion corrected averaging of all 16 phases. The template LV is segmented out using atlas-based joint label fusion [3] with diffeomorphic registration [4]. Finally, the template LV segmentation is warped back to the 16 phases using the previously derived inter-phase motion fields with corrective learning [5]. The workflow is summarized in Fig. 1.We conducted a leave-one-out cross validation study to evaluate the consistency between the automated and expert manual segmentations. The metrics used included Dice coefficient and mean minimum surface distance (MINDIST) of the two segmentations, and the correlation and agreement of the LV volumes derived from the automated and manual approaches.Fig. 2 shows an example segmentation result and the LV volumes throughout the entire cardiac cycle. Among all subjects and cardiac phases, the mean Dice coefficient and mean MINDIST were 0.93±0.03 and 0.13±0.08 mm, respectively. The LV volumes computed using the proposed method and manual segmentations had a correlation of 0.97 with an average error of 5.86% and showed excellent agreement, as shown in Fig. 3.In this work, we developed a fully automated method to segment LV from self-gated 4D MR images. The proposed technique showed excellent agreement with expert manual segmentations in our preliminary evaluation, and may potentially become a practical method for fast and accurate LV function analysis with minimum operator dependency. Future work will focus on performance tuning and further validations on both healthy and patient subjects.The proposed automated LV segmentation technique may enable “push-button” ventricular function assessment free of user intervention.