Explainable prediction of all-cause mortality from myocardial PET flow and perfusion images using deep learning

Introduction: Positron emission tomography (PET) myocardial perfusion imaging (MPI) provides complementary information regarding absolute myocardial blood flow (MBF) and relative perfusion, but it is challenging to optimally integrate this imaging information in clinical practice for risk assessment. We sought to develop and evaluate a novel explainable deep-learning (DL) network for the prediction of all-cause mortality (ACM) directly from PET MPI flow and perfusion polar map image data.

Methods: A total of 3,206 consecutive patients referred for regadenoson (91%) and adenosine (9%), stress and rest 82Rb PET were included. DL was trained using stress and rest polar map image data of raw perfusion, MBF, spill-over fraction and MFR combined with end-systolic and diastolic volumes, age, and sex. Patient data were randomly divided into 5 folds with repeated training, validation, and testing sets, using a 60%-20%-20% stratified split. The DL network comprised of 2 convolution blocks (convolution, batch normalization, dropout and maxpool layers) followed by 2 fully connected layers. The network was trained using binary cross-entropy loss function and optimized using a stochastic gradient-based optimization algorithm, Adam[ref 1]. The output of DL was a patient-specific ACM probability along with an attention map highlighting the polar map regions of importance for DL prediction. Stress total perfusion deficit (TPD), ischemic TPD (stress-rest TPD), rest/stress MBF and myocardial flow reserve (MFR), calculated with a 1-tissue compartment kinetic model were obtained with standard clinical software. The prediction of ACM by DL, standard perfusion and flow variables was evaluated using area under the receiver-operating-characteristic curve (AUC) and compared using Delong’s method.

Results: During median follow-up of 4.7 years (interquartile range, 3.4-6.0 years), 654 patients died (20.4%). The mean age was 70±12 years with 1,287 (40.1%) female patients. The average stress MBF was 2.7±1.1ml/g/min and MFR 2.5±1.0. The AUC for ACM prediction by DL (0.75 [0.73-0.77]) was significantly higher than stress TPD (0.63 [0.61-0.66]), ischemic TPD (0.62 [0.60-0.64]), stress MBF (0.66 [0.64-0.69]) or MFR (0.68 [0.65-0.70]) (p for all <0.0001). Patients in the highest quartile of risk by DL (DL score≥0.55) had an annual ACM rate of 11.8% (95% CI:10.56-13.2), with an 11.2-fold increase in the risk of death (95% CI:8.0-15.8) compared to the patients in the lowest quartile of risk (Dl score<0.12) who had annual ACM rate of 1.0% (95% CI:0.7-1.4). The DL network highlighted image regions associated with ACM risk directly on PET polar maps, along with a continuous probability of ACM risk, with per-patient results generated in less than 1 second.

Conclusions: The DL network was trained directly with PET dynamic flow and static perfusion PET polar map image data, which allows for improved patient risk stratification, compared to established quantitative methods for PET flow or perfusion assessment. The network also highlights specific polar map image regions associated with risk of mortality, providing an explanation of the patient-specific risk estimation to the physician.

• Download figure
• Open in new tab
• Download powerpoint