In the right direction: A deep learning tool for assessment of right ventricular function

Right Ventricular (RV) function is of significant prognostic importance in pulmonary hypertension (PH), as RV failure is the most common cause of death in this patient population.1, 2 Accurate and reproducible noninvasive measurements of RV function are critical in the management of patients with PH. Cardiac MRI (CMR) is considered the gold standard for the evaluation of RV function and RV ejection fraction (RVEF), but given limitations in scanner availability and patient tolerability, it can be challenging to perform CMR routinely in PH patients. As a result, the transthoracic echocardiogram (TTE) is the most commonly used modality for the assessment of RV size and systolic function.3 Quantitative assessment of the right ventricle by TTE is challenging due to the RV's complex shape. This is further exacerbated in patients with PH, in whom the RV often assumes a spherical morphology.4 Currently, RV free wall strain (FWS) and RV three-dimensional ejection fraction (RV 3DEF) are the most advanced clinically available methods for assessing RV function via TTE. They correlate well with CMR-derived RVEF and provide prognostic value in PH patients.5 Despite the development of auto-segmentation software for RV FWS and RV 3DEF, the widespread clinical adoption of these tools has been limited. Suboptimal visualization of the RV free wall is a common issue that may preclude the use of these methods in patients with PH. Moreover, many practitioners have not received training in these methods, which may lead to lower reliability and reproducibility of these tools in routine clinical practice.6 Visual assessment of 2-dimensional images, tricuspid annular lateral systolic velocity as assessed by tissue Doppler (S’), and M-mode-based tricuspid annular plane systolic excursion (TAPSE) remain the most widely used methods for assessment of RV function, despite their known limitations.6 As presented in this issue of Echocardiography,7 Murayama and colleagues have investigated a deep learning (DL) method to improve the assessment of RV function by TTE. As compared with prior studies that assessed different DL-based prediction models for RV EF by two dimensional (2D) TTE in healthy patient populations,8 this study enrolled patients with suspected or confirmed precapillary pulmonary hypertension. The DL model was tested and trained retrospectively on 85 examinations from 69 precapillary pulmonary hypertension patients to develop an algorithm to predict MRI-derived RVEF from 2D imaging. The authors found that their DL-predicted RVEF correlated with MRI derived RVEF. In receiver operating characteristic analysis, sonographer-derived fractional area change FAC and DL-predicted RVEF had high discriminatory power (area under the curve AUC .84 for DL RVEF, versus AUC .87 for FAC) for detection of severe RV dysfunction. We commend the authors for taking the initiative to move the evaluation of RV function beyond visual assessment, into the realm of quantitative analysis. Efforts to standardize the approach to RV assessment are critical, particularly because patients with early-stage disease may have subtle abnormalities in RV size and function that are difficult to detect if the assessment is incomplete. A gap remains between the capabilities of advanced echocardiographic techniques and utilization in actual clinical practice. Artificial intelligence (AI) in echocardiography interpretation has the potential to reduce cognitive errors and intra and interobserver variability. However, it is not evident that this DL tool improves upon previously validated RV assessment tools, such as RV 3D EF and RV FWS, as the authors did not perform a head-to-head comparison with these modalities. In our own echocardiography laboratory, performance of RV FWS has become standard for all TTEs with adequate acoustic windows across a range of patients, including those with PH. Addition of RV FWS with auto-segmentation software as part of the TTE evaluation has not added significantly to sonographer or echocardiographer workload and requires only minimal postprocessing by expert echocardiographers. In this small study, the time required to collect images for the DL model was not assessed. Additionally, several examinations had to be excluded from the DL model due to incomplete data due to poor acoustic windows and other challenges in image acquisition. AI-powered technology is playing an increasingly important role in health care. However, before this DL model or any another AI model could be introduced in clinical practice, the reported findings would need to be validated in real-world settings. In the meantime, implementation efforts should focus on incorporation of well-validated quantitative and semiquantitative techniques for RV assessment into routine clinical care. Ultimately, the most clinically useful echocardiographic analysis tools provide accurate and reproducible results, but just as importantly, they integrate seamlessly into sonographer and physician workflow.