Annotation and quality assessment of left ventricular filling and relaxation pattern using one-dimensional convolutional neural network

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Institute For Health Research (NIHR), UK Introduction Aberrations in left ventricular (LV) filling or relaxation – known as diastolic dysfunction – occur in heart failure with preserved ejection fraction. CMR is the reference modality for the assessment of ventricular systolic function, however, its role in evaluation of diastolic function is limited at present. One promising technique to assess diastolic function by CMR is the derivation of LV filling and emptying rates from the volume-time curves of cine images. Purpose To automatically assess the quality of LV filling-rate curves and annotate the peak emptying and filling rates. Methods A previously-described deep-learning network was used to automatically segment the entire cardiac cycle captured by short-axis SSFP cine images from the UK Biobank1. The LV filling-rate curves derived from the volume-time data were smoothed with Savitzky–Golay filter. The peak emptying rate (PER), early peak filling rate (PFR-E) and late peak filling rate (PFR-A) were first annotated by a simple peak finding algorithm from Python Scipy signal module. The preliminary annotated curves were reviewed by five human experts (i) to check for peak-annotation errors and (ii) to provide the curve quality score ranging from 1 to 3 for each peak (score 1 denotes good quality, score 2 represents moderate quality and score 3 indicates poor quality). Higher total score (minimum = 3, maximum = 9), therefore, represents poorer overall curve quality. This expert-annotated dataset was used to train two separate one-dimensional convolutional neural networks (1D-CNN) (Figure 1) for peak annotation and curve quality assessment (QA) using Tensorflow library in Python. Results The data from 6,328 LV filling-rate curves were split into the training and testing sets (80:20). The fine-tuned 1D-CNN comprising six hidden layers with two residual connections annotated the PER, PFR-E and PFR-A with the test-set accuracy of 95%, 95% and 98%, respectively. A second trained 1D-CNN for QA based on similar architecture predicted the overall curve quality score with a small error rate (mean absolute error: 0.46, mean squared error: 0.68). These two networks were used to quality check and label 19,409 UK Biobank CMR studies (See Figure 2 for exemplary results). After removing data from poor-quality curves (quality score ≥ 5), 18,735 studies remained. The mean±standard deviation of PER, PFR-E and PFR-A are 461±110 ml/s, 359±117 ml/s and 336±120 ml/s, respectively. Ageing is associated with lower PFR-E (−58.4 ml/s, 95% confidence interval [CI]: −56.1 to −60.7 ml/s per decade increment) and higher PFR-A (18.3 ml/s, 95% CI: 15.8 to 20.8 ml/s per decade increment). Conclusion The 1D-CNN models can be used to automatically grade the quality of LV filling rate curves and label important diastolic parameters with a high level of accuracy. The derived data recapitulate impaired LV relaxation pattern associated with ageing and can be used as surrogate indices of diastology by CMR. Figure 1Figure 2