Mean pulmonary artery pressure prediction with explainable multi-view cardiovascular magnetic resonance cine series deep learning model

Abstract

Background: Pulmonary hypertension (PH) is a heterogeneous condition and regardless of etiology impacts negatively on survival. Diagnosis of PH is based on hemodynamic parameters measured invasively at right heart catheterization (RHC); however, a non-invasive alternative would be clinically valuable. Our aim was to estimate RHC parameters non-invasively from cardiac magnetic resonance (MR) data using deep learning models and to identify key contributing imaging features.

Methods: We constructed an explainable convolutional neural network (CNN) taking cardiac MR cine series from four different views as input to predict mean pulmonary artery pressure (mPAP). The model was trained and evaluated on 1646 examinations. The model's attention weight and predictive performance associated with each frame, view, or phase were used to judge its importance. Additionally, the importance of each cardiac chamber was inferred by perturbing part of the input pixels.

Results: The model achieved a Pearson correlation coefficient of 0.80 and R² of 0.64 in predicting mPAP and identified the right ventricle region on short-axis view to be especially informative.

Conclusion: Hemodynamic parameters can be estimated non-invasively with a CNN, using MR cine series from four views, revealing key contributing features at the same time.

Keywords: Deep learning; Explainable AI; Mean pulmonary artery pressure; Multi-view cardiac MR; Pulmonary hypertension.

Graphical abstract

Fig. 1

An example of the processed CMR cine series at phase 1. (a)-(d) 2CH, 4CH, RVLA, and SAX views, respectivley. 2CH two-chamber, 4CH four-chamber, CMR cardiovascular magnetic resonance, RVLA right ventricular long-axis, SAX short axis

Fig. 2

Multi-view cine series model. (a) The model uses cine series from multiple views to generate a regression prediction. The input cine series are encoded into frame features, fused, and then processed by the Regression Layer to predict the mPAP. (b) The two-stage feature fusion strategy fuses frame features into intermediate features then into the final feature. (c) The Attention Feature Fusion Block (AFFB) is the key component in two-stage feature fusion. It fuses an arbitrary number of input features into one output feature of the same dimension by weighted-sum. 2CH two-chamber, 4CH four-chamber, mPAP mean pulmonary artery pressure, RVLA right ventricular long-axis, SAX short axis

Fig. 3

Scatter and Bland-Altman plots for predicting each hemodynamic parameter. Columns (a)-(c) are, respectively, the plots for the mPAP, PAWP, and PVR. mPAP mean pulmonary artery pressure, PAWP pulmonary artery wedge pressure, PVR pulmonary vascular resistance

Fig. 4

Confusion matrix of identifying precapillary PH in the absence of LHD non-invasively. LHD left heart disease, PH pulmonary hypertension

Fig. 5

Per-feature predictive performance (a) and attention weights (b and c). The round dots represent frame features, the squares represent intermediate features, and the triangle represents the final feature. Values within each sub-plots can be compared to establish relative importance—the higher the more important a feature is. Note that the values in (b) and (c) are not cross-comparable since they are never normalized together. 2CH two-chamber, 4CH four-chamber, mPAP mean pulmonary artery pressure, RVLA right ventricular long-axis, SAX short axis