High Resolution Isotropic 'Pseudo' 3D Cine imaging with Automated Segmentation using Concatenated 2D Real-time Imaging and Deep Learning.
TL;DR
The potential of creating a pseudo 3D-cine data from concatenated 2D real-time cine images using a series of DL models is demonstrated, which has short acquisition and reconstruction times with fully segmented data being available in less than one minute.
Abstract
Background
Conventional cardiovascular magnetic resonance (CMR) in pediatric and congenital heart disease uses 2D, breath-hold (BH), balanced steady state free precession (bSSFP) cine imaging for assessment of function, in addition to cardiac-gated, respiratory-navigated, static 3D bSSFP whole-heart imaging for anatomical assessment. Our aim is to concatenate a stack of 2D free-breathing real-time cines and use Deep Learning (DL) to create an isotropic fully segmented 'pseudo' 3D-cine dataset from these images.
Methods
Four DL models were trained on open-source data that performed: a) Interslice signal-correction; b) Interslice respiratory-correction; c) Super-resolution in the slice direction; and d) Segmentation of right and left atria and ventricles (RA, LA, RV, and LV), thoracic aorta (Ao) and pulmonary arteries (PA). Our method was validated in 20 patients undergoing routine cardiovascular examination, by converting prospectively acquired sagittal stacks of real-time cine images to segmented, isotropic pseudo 3D-cine data. Quantitative metrics (ventricular volumes and vessel diameters) and image quality of the DL pseudo-3D-cines were compared to reference-standard breath-hold cine and whole-heart imaging.
Results
All real-time data were successfully transformed into pseudo 3D-cines with a total offline reconstruction and post-processing time of <1min in all cases. There were no significant biases in any left ventricular (LV) or right ventricular (RV) metrics (bias ± standard deviation in ml, LV end diastolic volume (EDV): 0.7 ± 8.8, LV end systolic volume (ESV): -1.7 ± 7.1, RV EDV: 1.9 ± 12.4, RV ESV: -1.5 ± 10.7) with reasonable limits of agreement and correlation. There is also reasonable agreement for all vessel diameters, although there was a small but significant overestimation (p<0.05) of right PA (RPA) and main PA (MPA) diameter (RPA: bias = -1.0mm. MPA: bias = -1.1mm). The DL pseudo-3D-cine data were assessed to be of adequate diagnostic quality unlike the unprocessed 2D real-time data.
Conclusion
We have demonstrated the potential of creating a pseudo 3D-cine data from concatenated 2D real-time cine images using a series of DL models. Our method has short acquisition and reconstruction times with fully segmented data being available in less than one minute. Our models are trained from fully open-source datasets, allowing our technique to be easily shared with other clinical centers. The agreement with reference-standard imaging suggests that our method could help to significantly speed up CMR in clinical practice.