Deep learning for evaluating left atrium stress echocardiography: a proof of principle study

Introduction

Left atrial (LA) size is a dynamic variable that changes during stress echocardiography (SE) and provides valuable information within and beyond coronary artery disease. However, its measurement remains subjective, time-consuming and based on manual tracings.

Recent advances in deep learning might save analysis time and deflate variability by removing the subjectivity of LA assessment.

Purpose

In this proof of principle study we aim to validate the potential of an automated machine learning system for the evaluation of LA in SE.

Methods

From the image data bank of Stress Echo 2030 study, we selected 20 consecutive patients who underwent SE (using a variety of stress methods) in 2 recruiting centers. Imaging data was acquired in DICOM format and anonymized. The studies were reviewed by an expert cardiologist trained in SE evaluation who selected apical four chamber (A4Ch) view images during stress and rest phases and marked end-systolic (ES) and end-diastolic frames to be used in further evaluation. Endocardial borders for LA were traced in ES. The tracings were repeated by three different evaluators (expert cardiologist [C1], cardiologist from external center [C2] and a machine learning [ML] model – a convolutional neural network trained on an unrelated set of images). LA area (LAA), volume (LAV) using area-length formula and proportional LAV changes between stress and rest (ΔLAV = [rest − stress] / rest) were calculated. Each evaluator was blinded from each other's measurements.

Results

In total, 40 A4Ch images were acquired (20 at rest and 20 at stress) of which all were of sufficient quality for performing LA measurements using an automated system. Pearson correlation coefficients (R) for LAA were 0.95 (C1-ML), 0.96 (C2-ML) at rest and 0.88 (C1-ML), 0.79 (C2-ML) at stress. The C1-C2 pair had R of 0.98 and 0.86 for LAA at rest and stress.

LAV also showed good correlation between different raters with R values of 0.90 (C1-ML), 0.94 (C2-ML), 0.94 (C1-C2) at rest and 0.86 (C1-ML), 0.87 (C2-ML), 0.84 (C1-C2) at stress. Root mean squared errors (RMSEs) for LAV were 13.48 ml (C1-ML), 7.44 ml (C2-ML), 10.26 ml (C1-C2) at rest and 14.03 ml (C1-ML), 8.53 ml (C2-ML), 13.69 ml (C1-C2) at stress.

There were high level correlations between all raters for ΔLAV with R values of 0.93, 0.94 and 0.95 for C1-ML, C2-ML and C1-C2 pairs respectively. Mean (95% CI) ΔLAV values were −20.1 (−51.59 to +11.39), −14.95 (−43.77 to +13.87) and −14.37 (−34.46 to +5.72) for ML, C1 and C2. Comparison with one-way ANOVA did not show significant differences in mean ΔLAV values between operators (p=0.94).

Conclusions

Automated ML based system produces LA measurements that are comparable to human operators and can reduce the need for manual tracing. There was a tendency for all operators to have lower levels of agreement in stress images compared to rest, further showing the need for additional standardization of SE evaluation for machine and human operators.

Funding Acknowledgement

Type of funding sources: None.

De novo mutations in children born after medical assisted reproduction

STUDY QUESTION

Are there more de novo mutations (DNMs) present in the genomes of children born through medical assisted reproduction (MAR) compared to spontaneously conceived children?

SUMMARY ANSWER

In this pilot study, no statistically significant difference was observed in the number of DNMs observed in the genomes of MAR children versus spontaneously conceived children.

WHAT IS KNOWN ALREADY

DNMs are known to play a major role in sporadic disorders with reduced fitness such as severe developmental disorders, including intellectual disability and epilepsy. Advanced paternal age is known to place offspring at increased disease risk, amongst others by increasing the number of DNMs in their genome. There are very few studies reporting on the effect of MAR on the number of DNMs in the offspring, especially when male infertility is known to be affecting the potential fathers. With delayed parenthood an ongoing epidemiological trend in the 21st century, there are more children born from fathers of advanced age and more children born through MAR every day.

STUDY DESIGN, SIZE, DURATION

This observational pilot study was conducted from January 2015 to March 2019 in the tertiary care centre at Radboud University Medical Center. We included a total of 53 children and their respective parents, forming 49 trios (mother, father and child) and two quartets (mother, father and two siblings). One group of children was born after spontaneous conception (n = 18); a second group of children born after IVF (n = 17) and a third group of children born after ICSI combined with testicular sperm extraction (ICSI-TESE) (n = 18). In this pilot study, we also subdivided each group by paternal age, resulting in a subgroup of children born to younger fathers (<35 years of age at conception) and older fathers (>45 years of age at conception).

PARTICIPANTS/MATERIALS, SETTING, METHODS

Whole-genome sequencing (WGS) was performed on all parent-offspring trios to identify DNMs. For 34 of 53 trios/quartets, WGS was performed twice to independently detect and validate the presence of DNMs. Quality of WGS-based DNM calling was independently assessed by targeted Sanger sequencing.

MAIN RESULTS AND THE ROLE OF CHANCE

No significant differences were observed in the number of DNMs per child for the different methods of conception, independent of parental age at conception (multi-factorial ANOVA, f(2) = 0.17, P-value = 0.85). As expected, a clear paternal age effect was observed after adjusting for method of conception and maternal age at conception (multiple regression model, t = 5.636, P-value = 8.97 × 10-7), with on average 71 DNMs in the genomes of children born to young fathers (<35 years of age) and an average of 94 DNMs in the genomes of children born to older fathers (>45 years of age).

LIMITATIONS, REASONS FOR CAUTION

This is a pilot study and other small-scale studies have recently reported contrasting results. Larger unbiased studies are required to confirm or falsify these results.

WIDER IMPLICATIONS OF THE FINDINGS

This pilot study did not show an effect for the method of conception on the number of DNMs per genome in offspring. Given the role that DNMs play in disease risk, this negative result is good news for IVF and ICSI-TESE born children, if replicated in a larger cohort.

STUDY FUNDING/COMPETING INTEREST(S)

This research was funded by the Netherlands Organisation for Scientific Research (918-15-667) and by an Investigator Award in Science from the Wellcome Trust (209451). The authors have no conflicts of interest to declare.

TRIAL REGISTRATION NUMBER

N/A.

Automated evaluation of the right-sided heart function and geometry using deep learning

Funding Acknowledgements

Type of funding sources: None.

INTRODUCTION 

Deep learning (DL) has been successfully applied in the automated assessment of some transthoracic echocardiography (TTE) parameters such as left-ventricular ejection fraction. Nevertheless, automation of the right-sided heart assessment has not been widely studied, partially due to the relative difficulty involved in some of the right-sided heart measurement evaluation and time constraints in routine practice. Here we have explored the feasibility of a DL-based system capable of performing different tasks involved in the right-sided heart functional and geometric evaluation.

PURPOSE 

To develop a DL-based system assessing right atrium (RA) and right ventricle (RV) functional and geometric parameters and compare its accuracy to board-certified cardiologists.

METHODS

A total of 2,014 frames from 349 patients (with various indications for TTE) were used to train and validate four convolutional neural networks (CNNs) to perform either segmentation or landmark detection across four different TTE views: apical four-chamber (A4Ch), parasternal long-axis (PLAX), M-mode of tricuspid annulus and tissue Doppler imaging (TDI) of the right ventricular lateral wall. The CNNs were optimised to perform different right-sided heart measurements, namely, right atrial area in end-systole (RAA) and fractional area change (FAC) of RV in A4Ch view, proximal right ventricular outflow tract diameter (pRVOT) in PLAX view, tricuspid annular plane systolic excursion (TAPSE) in M-mode and S’ in TDI. Model performance was compared with two board-certified cardiologists using their average measurements on 20 test set patients. 

RESULTS

CNN predicted pRVOT diameter with a mean absolute error (MAE) of 1.02 mm and root mean squared error (RMSE) of 3.08 mm. The intersection over union (IoU) for the segmentation of RV and RA was 0.89 and 0.87, respectively. We then used  RV and RA segmentation predictions to calculate additional parameters which resulted in RMSE of 8.34% for FAC and 4.93cm2 for RAA. In the M-mode and TDI, the model achieved RMSE of 4.48 mm and 0.84 cm/s for the detection of TAPSE and S’, respectively.

CONCLUSIONS

We have demonstrated the feasibility of a DL-based system performing different measurements involved in right-sided heart evaluation. In a routine practice, where limited time resources might be available, such could assist in the thorough assessment of the right-sided heart geometry and function. Additional studies using cardiac magnetic resonance imaging to establish more precise accuracy of such systems is needed.

Accurate prediction of left ventricular diastolic dysfunction in 2D echocardiography using ensemble of deep convolutional neural networks

Introduction

Deep convolutional neural networks (CNNs) have been shown to be reliable in evaluating geometrically assessed systolic heart function, however, studies evaluating diastolic heart function, which reflects hemodynamic status, are scarce. This study presents a novel CNNs approach for recognition and evaluation of different diastolic left ventricular (LV) function measurements.

Purpose

To create a new CNNs approach in the assessment of LV diastolic function and compare the results with clinicians' measurements.

Methods

A total of 4586 2D echocardiographic images were extracted from the studies of 330 patients referred with varying indications. The ensemble of four CNNs was trained (80/20% training/validation patient split) for the assessment of transmitral inflow (E and A wave peak velocities), mitral septal and lateral annular velocities (e'), tricuspid regurgitation systolic velocity (Vmax) and detection the left atrial endocardium in apical four-chamber views. Additionally, E/A ratio, average E/e' ratio and left atrial volume index (LAVi) values were calculated for the evaluation of diastolic dysfunction according to the 2016 ASE/EACVI recommendations. CNNs performance in detecting diastolic dysfunction was compared to expert cardiologists on a set of 20 separate cases.

Results

Study results on the validation data showed that CNNs accurately predicted peak E/A ratio, average E/e' ratio (R2=0.88 and R2=0.89, respectively) and tricuspid regurgitation Vmax (R2=0.82). Figure 1 illustrates different functional LV measurements among the obtained echocardiographic images. Regarding the geometrical assessment of diastolic function, the segmentation model traced the left atrial endocardial border (intersection over union = 0.94) and subsequently was used in predicting the LAVi (R2=0.92). The ensemble of four CNNs had the area under the ROC curve of 0.93 for the detection of diastolic dysfunction when compared to expert cardiologists.

Conclusion

Deep CNNs demonstrate the capacity to detect peak velocities across different Doppler imaging modes and delineate left atrial endocardium border while using a relatively small dataset. Combining multiple CNNs has the potential of performing an accurate assessment of the diastolic LV function.

Funding Acknowledgement

Type of funding source: None

544 Deep learning in segmentation and function evaluation of right ventricle in 2D echocardiography

INTRODUCTION

Deep learning (DL) has been of increasing use in the field of echocardiographic cardiology. The importance of segmentation and recognition of different heart chambers was already presented in different studies. However, there are no studies made regarding the functional heart measurements. Even though, functional measurements of right ventricle (RV) remains "dark side of the moon", no doubtfully severity of RV dysfunction influences the worse outcomes.

PURPOSE

To evaluate DL for recognition of geometrical features of RV and measurement of RV fractional area change (FAC).

METHODS

A total of 896 end-systolic and end-diastolic frames from 129 patients (with various indications for the study) were used to train and validate the neural networks. Raw pixel data was extracted from EPIQ 7G (Philips) imaging platform. All of the images were from 2D echocardiography apical four chamber views. RV was annotated in each image, with 1716 images used for training and 180 for validation.

We used the state of art mask regional convolutional neural network (MR-CNN) and attention U-net convolutional neural network models for the RV segmentation task. Intersection over Union (IoU) was used as the primary metric for model evaluation. IoU measures the number of pixels common between the target and the prediction masks divided by the total number of pixels present across both masks. Additionally FAC was calculated using frames with minimal and maximal segmented area by the network.

RESULTS

U-Net architecture demonstrated considerably faster training compared to MR-CNN with time per training step of 85 ms and 750 ms for U-Net and MR-CNN, respectively (p &lt; 0.001). MR-CNN and U-Net had an IoU of 0.91 and 0.89 respectively on validation dataset which corresponds to good performance of the model and there was no significant difference between the different neural networks (p = 0.876). Comparing the evaluation of FAC by physician and U-Net the mean squared difference was 12% when using minimum and maximum right ventricle area detected by the network.

CONCLUSION

With small dataset deep learning give us ability to recognize RV and measure RV FAC in apical four chamber view with high accuracy. This method offers assessment of RV to become daily used in the cardiologist practice, moreover, in the near future automated measurements will allow to reduce the need of observer manual evaluation. Improvements can be made in FAC calculation by also improving techniques for end-systolic and end-diastolic frame detection.

P1465Artificial intelligence in echocardiography - Steps to automatic cardiac measurements in routine practice

Introduction

The growth of artificial intelligence (AI) use in echocardiography over the past years has been exponential, proposing new paths to overcome inter-operator variability and experience of the operator. Even though the applications of AI are still in their infancy within the field of echocardiography, the potential of AI implies future directions and is eager to assist for accuracy and efficiency of manual tracings. Deep learning, a subset of machine learning algorithms, is gaining popularity in echocardiography as a state of the art in visual data analysis.

Purpose

To evaluate deep learning for two initial tasks in automated cardiac measurements: view recognition and end-systolic (ES) and end-diastolic (ED) frame detection.

Methods

A total of 230 patients' (with various indications for study) 2D echocardiography data was used to train and validate neural networks. Raw pixel data was extracted from EPIQ 7G, Vivid E95 and Vivid 7 imaging platforms. Images were labeled according to their view: parasternal long axis (PLA), basal short axis, short axis at mitral level, apical two, three and four chambers (A4C). Additionally, ES and ED frames were labeled for A4C and PLA views. Images were de-identified by applying black pixel masks to non-anatomical data and removing metadata. Convolutional Neural Network (CNN) architecture was used for the classification of 6 different views. A total of 34752 and 3972 (5792 and 662 per view) frames were used to train and validate the network, respectively. Long-term Recurrent Convolutional Network (LRCN) combining temporal and spatial cognition was used for ES and ED frame detection. A total of 195 and 35 sequences with a length of 92 frames were used to train the LRCN.

Results

CNN for view classification had an AUC of 0.95 (sensitivity 95%, specificity 97%). Accuracy was lower for visually similar views, namely apical three-chamber and apical two-chamber. Training for ES and ED detection was achieved when training LRCN for regression instead of classification of each frame. LRCN for cardiac cycle evaluation had an average Framed Difference (aFD) of 2.31 (SD±2.15) for ED and 1.97 (SD±2.04) frames for ES detection which corresponds to error rate of about 0.04 s.

Conclusion

Determining echocardiographic view and evaluating cardiac cycle are the first steps in automating cardiac measurements. We have demonstrated the potential of two deep learning algorithms in accomplishing these tasks. Initial results are promising for the development of neural networks for cardiac segmentation and measuring of anatomical structures.