Predicting cardiovascular risk traits from pet myocardial perfusion imaging with deep learning

Abstract

Abstract Background and aim Deep Learning has revolutionised image analysis and its implementation in cardiology is rapidly advancing. DL offers the possibility to abstract highly-complex patterns from any image in order to optimise classification and prediction tasks. At the same time, positron emission tomography (PET) represents the reference technique for quantitative evaluation of myocardial perfusion and is a powerful tool for the diagnosis of myocardial ischemia and infarction. Although pathological perfusion patterns in PET are easily recognised by expert clinicians, it is unclear whether PET images harbour complex patterns inherent to traditional cardiovascular risk traits (factors) recognisable in an individual-patient basis. Hence, we aimed to deploy Deep Learning in order to predict cardiovascular risk traits from individual quantitative PET myocardial perfusion images. Methods Data form 1180 patients evaluated through quantitative N13-ammonia PET-imaging for suspected myocardial ischemia was analysed. We implemented transfer learning with fine-tuning of individual concatenated instances of an ImageNet pre-trained convolutional neural network (ResNet-50). From each PET scan, the 3 standard quantitative polar maps (rest/stress/perfusion reserve) were used as network input. A 5-fold cross validation policy was applied to training and validation for hyperparameters optimisation. Deep Learning modelling was deployed to independently predict: sex, smoking, arterial hypertension, dyslipidemia and type 2 diabetes mellitus as cardiovascular risk traits. Final model performance was evaluated through AUC and accuracy (with associated standard deviations [SD]) on a hold-out test set of 104 scans. Deep Learning-based heat maps were generated to identify regions of interest in PET imaging related to each of the predicted risk trait Results Deep Learning was able to strongly predict sex (0.87±0.04, 78%±4) and diabetes mellitus (0.75±0.13,71%±8), while its performance was only discrete for smoking (0.63±0.11, 85%±2), hypertension (0.61±0.06,58%±3) and dyslipidemia (0.59±0.05,57%±3), albeit all statistically significant (p<0.01). Conclusion Deep Learning is able to significantly predict cardiovascular risk traits from individual quantitative PET myocardial perfusion images. This suggests the existence of complex high-dimensional and localised features within cardiac imaging that relate to cardiovascular risk factors at the individual level, which definitely warrants further research. Funding Acknowledgement Type of funding source: None