Dynamic frame-by-frame motion correction for 18F-flurpiridaz PET-MPI using convolution neural network

Abstract

Purpose: Precise quantification of myocardial blood flow (MBF) and flow reserve (MFR) in ¹⁸F-flurpiridaz PET significantly relies on motion correction (MC). However, the manual frame-by-frame correction leads to significant inter-observer variability, time-consuming, and requires significant experience. We propose a deep learning (DL) framework for automatic MC of ¹⁸F-flurpiridaz PET.

Methods: The method employs a 3D ResNet based architecture that takes 3D PET volumes and outputs motion vectors. It was validated using 5-fold cross-validation on data from 32 sites of a Phase III clinical trial (NCT01347710). Manual corrections from two experienced operators served as ground truth, and data augmentation using simulated vectors enhanced training robustness. The study compared the DL approach to both manual and standard non-AI automatic MC methods, assessing agreement and diagnostic accuracy using minimal segmental MBF and MFR.

Results: The area under the receiver operating characteristic curves (AUC) for significant CAD were comparable between DL-MC MBF, manual-MC MBF from Operators (AUC=0.897, 0.892 and 0.889, respectively; p>0.05), standard non-AI automatic MC (AUC=0.877; p>0.05) and significantly higher than No-MC (AUC=0.835; p<0.05). Similar findings were observed with MFR. The 95% confidence limits for agreement with the operator were ±0.49ml/g/min (mean difference = 0.00) for MFR and ±0.24ml/g/min (mean difference = 0.00) for MBF.

Conclusion: DL-MC is significantly faster but diagnostically comparable to manual-MC. The quantitative results obtained with DL-MC for MBF and MFR are in excellent agreement with those manually corrected by experienced operators compared to standard non-AI automatic MC in patients undergoing ¹⁸F-flurpiridaz PET-MPI.

Keywords: 18F-flurpiridaz; PET; deep learning; motion correction; myocardial blood flow; myocardial flow reserve.

FIG. 1

Proposed architecture for motion translation vector prediction. Input channel 1 represents the given frame (m) and channel 2 is the reference frame, intensity summed frames with 120sx: translation x direction, t_y: translation y direction, t_z: translation z direction, f(m): function of frame number m

FIG. 2

Bootstrap probability density distributions of the original manual motion displacement corrections per frame (in mm) of the dynamic PET scans in X (red), Y (green) and Z (blue) directions. Simulated motion vectors that emulate real patient motion are derived from these distributions for data augmentation

FIG. 3

Site wise splitting of dataset to achieve five-fold external validation

FIG. 4

Bland-Altman and Correlation plots of (a) Standard non-AI automatic motion correction (MC) vs. average (Avg) manual MC and (b) Deep Learning (DL)-based MC vs. average manual MC for myocardial flow reserve (MFR). The top row depicts Bland–Altman plot and the bottom row shows the correlation plots. CCC: Concordance correlation coefficient, CI: Confidence interval

FIG. 5

Bland-Altman and Correlation plots of (a) Standard non-AI automatic motion correction (MC) vs. average (Avg) manual MC and (b) Deep Learning (DL)-based MC vs. average manual MC for stress and rest minimal segmental myocardial blood flow (MBF). The top row depicts Bland–Altman plot and the bottom row shows the correlation plots. CCC: Concordance correlation coefficient, CI: Confidence interval

FIG. 6

Concordance Correlation Coefficients (CCC) of Deep Learning (DL)-based motion correction (MC) and average (Avg) manual MC for (a) myocardial flow reserve (MFR) and (c) myocardial blood flow (MBF), per segment. Per segment CCC for (b) MFR and (d) MBF with manual motion corrections from two different operators

FIG. 7

Per-patient diagnostic performance of minimal segmental (a) stress myocardial blood flow (MBF), (b) myocardial flow reserve (MFR) with residual activity correction applied, compared among manual motion correction (MC) from Operator 1 (O1) (light blue), Operator 2 (O2) (dark blue), standard non-AI automatic MC (orange), deep learning (DL) based automatic MC (red) and no-MC (grey). AUC: area under the receiver operating characteristic curve; CI: confidence interval; *: significant (p < 0.05)