Deep Learning-based Measurement of Isovolumic Relaxation Time from Cardiovascular Magnetic Resonance Long-axis Cines: Validation with Pressure-derived IVRT

Isovolumic relaxation time (IVRT) is an important timing metric of diastolic dysfunction (1), but is not commonly used and is not available for evaluation by cardiovascular magnetic resonance (CMR). New deep learning methods (2-4) can automatically track the mitral and tricuspid valve plane motion on long-axis cines and potentially automate IVRT measurement.  We propose and evaluate a new method for deriving IVRT in the left (LV) and right ventricles (RV) by automated measurement of the systolic quiescent time.

Methods:

This study included three cohorts: 67 patients  for validation of automated IVRT against manual measurement of LV- and RV-IVRT;  subjects with pulmonary hypertension (n =22,14 female, 62 ± 17 years old) who had right heart catheterization (RHC) within two weeks of CMR, for comparing pressures with IVRT;   and healthy subjects  for measuring LV-IVRT (N=43, 25 females, 61 ± 11 years old) and RV-IVRT (n=18, 9 females, 50 ± 10 years old).  Long-axis bSSFP cines (34ms true frame-times) were annotated using previously-trained deep-learning networks MVnet and TVnet (2,3) to identify valve insertion points, throughout the cardiac cycle.  In 67 subjects, cines were also analyzed manually with annotation of mitral and tricuspid valve insertion points, using Segment software (Medviso). For both automated and manual annotations, IVRTs were calculated as the systolic time-period when the valve plane was within 5% of maximal systolic excursion as shown in Figure 1. In patients with pulmonary hypertension and recent RHC, correlation studies were performed between invasive pressure-derived IVRT and automated cine-based IVRT.  To estimate IVRT from RV pressure curves, RV-IVRT was defined using the 2^nd derivative of the RV pressure curves (Figure 2A), as described by others (5). LV-IVRT was measured in this cohort, and compared to pulmonary capillary wedge pressure (PCWP).

Results:

In healthy subjects, LV-IVRT was 106± 26ms, while RV-IVRT was 92± 23ms, using automated IVRT measurement.  Automated IVRT showed very strong and strong correlations with manual annotations for both LV-IVRT (ICC=0.84) and RV-IVRT (ICC=0.79), respectively in 67 subjects. The mean bias (±2 standard deviation) expressed in % of the RR were -0.41 ± 6.3 % and 0.00 ± 5.1 % for LV IVRT and RV-IVRT.   In patients with pulmonary hypertension (N=22), cine-based RV-IVRT showed a strong positive correlation with pressure-derived RV-IVRT (R=0.63) as shown in Figure 2B, which is expected and serves to partially validate cine-based RV-IVRT.   LV-IVRT demonstrated moderate negative correlation with invasively measured filling pressure PWCP in patients with pulmonary hypertension (R=-0.51), as expected,  and described previously (6). 

Conclusion:

Automated valve-tracking-based estimation of IVRT was validated (vs. manual approach) and correlates to pressure-based measurements of IVRT. Automated IVRT may have clinical utility for evaluating diastolic dysfunction.

 

Figure 1: Measurement of IVRT based on manual or automatic tracking of the mitral valve. A) Examples of mitral valve point annotation and valve-plane tracking. B) Plot of mitral valve excursion using 2 and 4ch-cines, with LV-IVRT defined as the systolic quiescent time during which the mitral valve is within 5% of MAPSE (mitral annular plane systolic excursion). C) Similarly, the RV-IVRT is defined (using 4ch cines only), based on 5% of TAPSE (tricuspid annular plane systolic excursion).

Figure 2: A) RV pressure curves in a pulmonary hypertension patient. RV IVRT was determined using the 2nd derivative of the pressure curve to determine the beginning and end points of isvolumic relaxation, as shown, averaged over three cycles. B) Comparison of RV-IVRT by automated cine analysis vs. RV pressure curves, showing substantial agreement.