A machine learning cardiac magnetic resonance approach to extract disease features and automate pulmonary arterial hypertension diagnosis

Abstract

Aims: Pulmonary arterial hypertension (PAH) is a progressive condition with high mortality. Quantitative cardiovascular magnetic resonance (CMR) imaging metrics in PAH target individual cardiac structures and have diagnostic and prognostic utility but are challenging to acquire. The primary aim of this study was to develop and test a tensor-based machine learning approach to holistically identify diagnostic features in PAH using CMR, and secondarily, visualize and interpret key discriminative features associated with PAH.

Methods and results: Consecutive treatment naive patients with PAH or no evidence of pulmonary hypertension (PH), undergoing CMR and right heart catheterization within 48 h, were identified from the ASPIRE registry. A tensor-based machine learning approach, multilinear subspace learning, was developed and the diagnostic accuracy of this approach was compared with standard CMR measurements. Two hundred and twenty patients were identified: 150 with PAH and 70 with no PH. The diagnostic accuracy of the approach was high as assessed by area under the curve at receiver operating characteristic analysis (P < 0.001): 0.92 for PAH, slightly higher than standard CMR metrics. Moreover, establishing the diagnosis using the approach was less time-consuming, being achieved within 10 s. Learnt features were visualized in feature maps with correspondence to cardiac phases, confirming known and also identifying potentially new diagnostic features in PAH.

Conclusion: A tensor-based machine learning approach has been developed and applied to CMR. High diagnostic accuracy has been shown for PAH diagnosis and new learnt features were visualized with diagnostic potential.

Keywords: diagnosis; machine learning; pulmonary arterial hypertension; right ventricle; tensor.

Figure 1

Illustration of multilinear principal component analysis (MPCA). PCA is a traditional linear dimensionality reduction method that extracts low-dimensional features from high-dimensional input. MPCA extends PCA to tensor representations of data. In this article, the input to MPCA is N samples of CMR sequences with size I × J × K, with a spatial dimension of I × J and a time dimension of K (frames). MPCA maps each I × J × K tensor to a low-dimensional P × Q × R tensor (P < I, Q < J, and R < K) using three projection matrices of size I × P, J × Q, and K × R. During training, these three matrices are optimized to maximize the variation captured in the N mapped P × Q × R tensors and these optimized matrices are the output of the MPCA learning algorithm. During testing, the learnt three matrices map a new I × J × K tensor input of I × J × K into a P × Q × R tensor as its low-dimensional representation.

Figure 2

Proposed machine learning workflow: manual landmarking, pre-processing steps with registration, masking, and rescaling, machine learning steps, visualization of learnt factors and features, and feature maps. The learnt factors in the fourth row are highly interpretable: the column (size I × 1) and row (size J × 1) factors capture spatial variations closely resemble wavelets popular in representing fundamental patterns in natural images, and the time factor (size K × 1) shows the cardiac phase.

Figure 3

Patient flow diagram.

Figure 4

Representative feature map images: (i) short-axis (A) and four-chamber (B) CMR images in a patient with PAH, taken from Phase 11, early diastole. Notably, features are identified at the level of the interventricular septum and close to the right ventricular outflow tract within the blood pool. (ii) Short-axis (C) and four-chamber (D) CMR images in a patient without PAH, taken from Phase 1, end-diastole. Features are noted at the level of the mid chamber/apex septum and basal LV lateral wall. Red features are more indicative of PAH and green indicative of no PAH. The colour bars show the amplitude of the features detected, with a higher absolution value indicating a higher importance.

Figure 5

Receiver operating characteristic curve analysis showing diagnostic accuracy of the proposed machine learning approach in identification of PAH and IPAH, using short-axis (scale: 64 × 64) and four-chamber images (scale: 128 × 128) and small ellipse mask.