Deep Learning for Explainable Estimation of Mortality Risk From Myocardial Positron Emission Tomography Images

Abstract

Background: We aim to develop an explainable deep learning (DL) network for the prediction of all-cause mortality directly from positron emission tomography myocardial perfusion imaging flow and perfusion polar map data and evaluate it using prospective testing. Methods: A total of 4735 consecutive patients referred for stress and rest 82 Rb positron emission tomography between 2010 and 2018 were followed up for all-cause mortality for 4.15 (2.24–6.3) years. DL network utilized polar maps of stress and rest perfusion, myocardial blood flow, myocardial flow reserve, and spill-over fraction combined with cardiac volumes, singular indices, and sex. Patients scanned from 2010 to 2016 were used for training and validation. The network was tested in a set of 1135 patients scanned from 2017 to 2018 to simulate prospective clinical implementation. Results: In prospective testing, the area under the receiver operating characteristic curve for all-cause mortality prediction by DL (0.82 [95% CI, 0.77–0.86]) was higher than ischemia (0.60 [95% CI, 0.54–0.66]; P <0.001), myocardial flow reserve (0.70 [95% CI, 0.64–0.76], P <0.001) or a comprehensive logistic regression model (0.75 [95% CI, 0.69–0.80], P <0.05). The highest quartile of patients by DL had an annual all-cause mortality rate of 11.87% and had a 16.8 ([95% CI, 6.12%–46.3%]; P <0.001)-fold increase in the risk of death compared with the lowest quartile patients. DL showed a 21.6% overall reclassification improvement as compared with established measures of ischemia. Conclusions: The DL model trained directly on polar maps allows improved patient risk stratification in comparison with established methods for positron emission tomography flow or perfusion assessments.

Deep Learning and Artificial Intelligence: What Does the Cardiologist Really Need to Know?

“…Singh et al 7 have presented an interesting a practical application of deep learning to integrate image data with clinical information to enhance interpretive power. As this and other techniques find their way into everyday use, it is essential we understand the strengths and weaknesses of these methods.…”

“…Singh et al 7 present an interesting new addition to this family of deep learning models. The model presented combines image data with clinical data in much the same way as human clinician would do.…”

Explainable Artificial Intelligence and Cardiac Imaging: Toward More Interpretable Models

“…For instance, Pérez-Pelegrıó et al 28 developed a new explainable approach that combines class activation mapping with U-net to automatically estimate the LV volume in end diastole and obtain the result in the form of a segmentation mask without segmentation labels to train the algorithm. Grad-CAM was used in 7 cardiac imaging studies, either for classification 18,34,40,[48][49][50] or segmentation. 51 The latter in particular proposed a new interpretable CNN model (fast and accurate echocardiographic automatic segmentation based on U-Net) that integrates U-net architecture and transfer learning (from Visual Geometry Group 19) to segment 2-dimensional echocardiography of 88 patients into 3 regions (LV, interventricular septal, and posterior LV wall).…”

“…In addition, 4 studies relied on SHAP to interpret the model outputs. [32][33][34][35] One of these studies 32 relied on SHAP to develop and test an explainable ML model to assess whether noncontrast CMR could bring added value to predicting HF hospitalization compared with clinical data only. Their results demonstrated that CMRbased ML models could provide a significantly superior prediction of HF hospitalization (area under the curve, 0.81) compared with the basic clinical model (area under the curve, 0.64).…”

“…32 SHAP and Grad-CAM were also integrated in a single study to build an XAI model using myocardial positron emission tomography images to predict mortality risk. 34 Grad-CAM was used to create attention maps in the model, whereas SHAP was incorporated to determine the importance of input.…”