Towards deep learning methods for quantification of the right ventricle using 2D echocardiography

1. Background

Right ventricular (RV) dysfunction can be found in up to 20% of people who undergo cardiac evaluation and has been independently associated with increased mortality [1–3]. Reliable, noninvasive evaluation of RV size and function evaluation is required for the diagnosis and treatment of patients with RV dysfunction; however, there is a large gap between available imaging modalities [4].

Today, cardiac magnetic resonance imaging (CMR) is the optimal modality for RV assessment, allowing for precise and quantitative analysis [5]. However, expensive equipment, specialized training, and high costs still limit the wide use of CMR [6].

Two-dimensional echocardiography (2DE) is a first-line modality for RV assessment but suffers from decreased accuracy and its measurements are not easily reproducible. Moreover, 2DE does not allow for quantitative analysis of RV volume ejection fraction (RVEF) [4,7]. Prior 2DE quantitative methods that relied on geometric assumptions to calculate RV volumes were not successful as a result of the complexity of the RV shape [8]. The low accuracy of 2DE in RV evaluation has significant implications as patients with RV dilatation and/or dysfunction may not be diagnosed, which in turn may delay their treatment. In their quality control study, Ling et al. analyzed the evaluation of 12 patients by 15 different readers and compared their results to CMR. The analysis showed that 40% of patients with RV dilatation and 30% of patients with low RVEF were missed [9].

2. Deep learning for 2DE RV evaluation

Deep learning (DL) has revolutionized cardiovascular imaging [10–12] and may prove useful in bridging the gap between CMR and 2DE in RV evaluation. Prior attempts to use DL for this purpose have been limited so far. A study by Knight et al. in 2015 attempted to use tracings from seven 2DE views in order to calculate RV volumes [13]. The proposed method used an online database of CMR images as a reference for calculations. Although the results were accurate and reproducible, the method did not achieve wide acceptance, likely because it is time-consuming and tedious [13]. Recently, a study by Tokodi et al. used a DL algorithm to predict RVEF from 2DE [14]. The authors used a large dataset to create a fully automated DL method that is based on a single 2DE RV view. Results were compared with visual assessment of expert cardiologists. Despite the fact that the study was promising, the achieved R² was only 0.45, which means that the majority of RVEF variance was not explained. In addition, the sensitivity for detection of decreased RVEF was 0.727, therefore missing more than 25% of patients with RV dysfunction [15]. Although the authors are to be praised, the results of the DL method reflect the limitation of a single view to evaluate the RV.

3. A novel DL approach for quantitative RV evaluation

An initial usage of multiple 2DE RV views would allow a better assessment of the 3D RV shape, as even Simpson's rule for left ventricular ejection fraction requires 2 views. Our group recently proposed a DL pipeline to harness the DL potential for quantitative 2DE RV evaluation [16].

The method involves tracing the RV myocardium in 8 standardized views (parasternal long axis [PLAX], RV-Inflow, parasternal short axis at the level of the aortic valve [PSAX-AV], base [PSAX-base], mid [PSAX-mid] and apex of the left ventricle [PSAX-distal], standard four-chamber [Four-C] and subcostal four-chamber [Sub-C]). A focused apical RV view was used when the standard Four-C view was not available or had poor image quality. The area for each tracing is calculated and then used by a Feature Tokenizer and Transformer (FTT) algorithm [17] for the prediction of CMR-derived RV end-diastolic volume (RVEDV) and RV end-systolic volume (RVESV). The patient's age and gender are also used for this prediction. RVEF is then calculated from RV volumes.

We applied this method to 50 adult patients who had 2DE and CMR performed within a 30-day period [18]. For all patients, both CMR and 2DE were performed according to the judgment of the treating cardiologist. Patients with prior cardiac surgery of the RV, complex congenital heart disease, more than small pericardial effusion, cardiac tamponade, and atrial fibrillation at the time of either study or severe RV dysfunction were excluded. We excluded patients with inadequate imaging quality. The median age of the included patients was 51 years (interquartile range 32–62) and 42% were women. Symptoms of heart failure were present in 19 patients (40%). The most common underlying diseases included coronary artery disease (22%), hypertrophic cardiomyopathy (14%) and simple congenital defects (12%). RVEDV, RVESV, and RVEF were 163 ± 70 ml, 82 ± 42 ml, and 51 ± 8%, respectively, by CMR. RV dilatation was present in six (12%) patients and RV dysfunction in 9 (18%), when using published CMR criteria [19]. Patients were randomly split into training, validation, and testing subsets (35, 5, and 10 patients, respectively).

The FTT achieved good accuracy for RV volume (R² = 0.953, absolute percentage error [APE] = 9.75 ± 6.23%) and RVEF (APE = 7.24 ± 4.55%). A Bland-Altman analysis also showed good agreement between FTT and CMR for RVEDV (1.27 ± 23.35 ml), RVESV (-2.61 ± 19.63 ml) and RVEF (-1.97 ± 7.04%). In addition, the FTT showed 100% diagnostic accuracy for the detection of RV dilatation and 90% for the detection of RV dysfunction, when compared with CMR. Reliability analysis was also performed, showing poor to moderate intraobserver reliability for PSAX views of the RV at the mid and apical level and moderate interobserver reliability for mid-PSAX views. Measurements in all other views showed good to excellent intra- and inter-observer reliability.

The prior results could be considered the initial step toward a new, DL-based method for accurate 2DE evaluation of RV. According to this method, a cardiologist would perform planimetry of the RV from different views in a way similar to Simpson's rule for the left ventricle. However, the use of 8 views represents an important barrier to the clinical application of this DL method. To address this, we assessed in a separate study whether the number of views can be reduced without sacrificing diagnostic accuracy. An explainability analysis using the Instance-Based Uncertainty Quantification (IBUG) method, and a CatBoost classifier [20] showed that the 3 most important 2DE RV views for the prediction of RV volumes are the following: PLAX, Four-C, and PSAX-base. These results are consistent with the reliability analysis, which ensured good reliability for these three views.

4. Toward a new method for accurate 2DE RV evaluation using DL

The diagnostic accuracy of the proposed FTT method using a reduced number of 2DE views remains to be examined. Specifically, the diagnostic accuracy of each combination of the previously mentioned three views has to be analyzed. The resulting method could make use of 1-3 2DE views for RV quantification. If there is a significant decrease in the diagnostic accuracy of the FTT method with reduced RV views, then additional geometric features (e.g., eccentricity, long and short axis) of the planimetered shapes could be analyzed.

Following that, the FTT method will have to be validated in a larger dataset of patients with broader RV size and function. If successful, such a method would allow for accurate, quantitative, and widely available RV evaluation using established ultrasound technology, bridging the current gap between 2DE and CMR. Such a pipeline could additionally function as a gatekeeper for CMR.

5. Conclusion

DL is revolutionizing cardiovascular imaging and may prove extremely useful in bridging the large gap between 2DE and CMR in the evaluation of RV. Herein, we discussed the current status of a proposed DL-based method that shows accuracy similar to CMR. We further outlined further steps in order to reach a method that resembles the established Simpson's rule for the left ventricle.

Funding Statement

T Goldstein supported this work with a private donation.

Financial disclosure

T Goldstein supported this work with a private donation. The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Competing interests disclosure

The authors disclose that they have applied for a provisional patent for their related work which is referenced in this letter. The authors have no other competing interests or relevant affiliations with any organization or entity with the subject matter or materials discussed in the manuscript apart from those disclosed.

Writing disclosure

No writing assistance was utilized in the production of this manuscript.