Right ventricular (RV) function is a major determinant of clinical outcomes in patients undergoing cardiac surgery and transcatheter interventions. Although preprocedural assessment has traditionally focused on left ventricular (LV) morphology and function, RV dysfunction is now recognized as a key predictor of morbidity and mortality across a wide range of cardiovascular diseases. This review summarizes the current literature on the prognostic significance of RV function in valvular heart disease in the setting of surgical or interventional procedures. In valvular disorders, assessment of RV function is of particular importance, as adaptation to chronic pressure and volume overload, followed by the onset of decompensation, often precedes manifest clinical deterioration. This applies not only to tricuspid valve disease but also to mitral and aortic valve pathologies, in which secondary pulmonary hypertension and functional tricuspid regurgitation play a central role in RV remodeling and functional decline. Conventional two-dimensional echocardiographic parameters, such as tricuspid annular plane systolic excursion (TAPSE), RV fractional area change (FAC), and tissue Doppler–derived systolic peak velocity (TDI S?), provide limited information due to the complex geometry of the RV. Contemporary imaging modalities, including speckle-tracking echocardiography, three-dimensional echocardiography, and cardiac magnetic resonance imaging, enable more accurate quantification of RV volumes, ejection fraction, strain parameters, RV motion components, and right ventricle–pulmonary artery coupling, and demonstrate superior prognostic value compared with traditional methods. Accordingly, comprehensive evaluation of RV function using modern imaging techniques should be incorporated into routine clinical practice to optimize risk stratification and preprocedural planning prior to cardiac surgery and structural interventional procedures.

Background: Transcatheter tricuspid valve intervention (TTVI) has emerged as a valuable therapeutic option for patients with severe tricuspid regurgitation. However, the impact of TTVI on right ventricular (RV) function remains incompletely understood, partly due to the limitations of conventional echocardiographic parameters. Objectives: The purpose of this study was to evaluate RV functional trajectories in patients undergoing TTVI using a deep learning model that estimates RV ejection fraction (RVEF) from two-dimensional apical four-chamber view echocardiographic videos. Methods: This single-center analysis included 373 patients undergoing TTVI for severe tricuspid regurgitation between 2018 and 2023. A previously published and thoroughly validated deep learning model was used to predict RVEF at baseline and 1 to 3 days after the procedure. The primary endpoint was 1-year all-cause mortality. Results: Although the median deep learning–predicted RVEFs were similar before and after TTVI at the cohort level, individual trajectories diverged. Using maximally selected log-rank statistics, an optimal prognostic threshold of 38% for postprocedural RVEF was identified. Patients below this threshold showed significantly worse 1-year survival compared to those above it (58.4% vs 85.1%; HR: 3.12; P < 0.001). RVEF in this high-risk group had declined from 41% (IQR: 38%-44%) at baseline to 36% (IQR: 35%-37%) postprocedurally (P < 0.001). Conclusions: Deep learning enabled an unbiased echocardiographic assessment of RV function after TTVI and identified a high-risk group with poor outcomes. These findings are exploratory and require external validation; if confirmed, deep learning–enhanced echocardiography may improve risk stratification and guide personalized follow-up strategies in patients undergoing TTVI.

Intense exercise imposes hemodynamic load on the heart, and while morphological remodeling is well-characterized, assessment of exercise-induced functional changes, like enhanced contractility, remains challenging. We aimed to introduce a novel 3D echocardiography (3DE)-derived method for noninvasive quantification of biventricular systolic function, less dependent on loading conditions in competitive athletes, and to explore the relationship with peak exercise capacity. We enrolled 260 athletes and 24 sedentary volunteers. All subjects underwent 3DE to measure left (LV) and right ventricular (RV) volumes and ejection fractions (EF). Biventricular global longitudinal strain (GLS) tracings and noninvasively estimated pressure curves were concatenated and further adjusted to instantaneous volumes to create pressure-strain-volume loops and derive volume-adjusted myocardial work (MW) indices (LV GWIV and RV GWIV). Athletes had lower biventricular EF and LV GLS, but significantly higher LV GWIV (10273 ± 2929 vs. 7387 ± 2050 mmHg%·mL, p < 0.001) and RV GWIV (3422 ± 1339 vs. 2436 ± 796 mmHg%·mL, p < 0.001) compared to controls. Among the functional echocardiographic parameters, RV GWIV showed the strongest correlation (r = 0.30, p < 0.001) and was an independent predictor of exercise capacity. Our novel metrics captured enhanced biventricular function in athletes at rest, and RV GWIV was independently associated with higher peak exercise capacity.

Right ventricular (RV) function is a major determinant of clinical outcomes in patients undergoing cardiac surgery and transcatheter interventions. Although left ventricular morphology and function have been the traditional focus of preprocedural assessment, RV dysfunction is now recognized as an important predictor of morbidity and mortality. Therefore, inclusion of RV-related parameters in preprocedural risk assessment is on the rise. This review summarizes current evidence on the role of RV function in various cardiac diseases that require surgery or interventions. Conventional two-dimensional echocardiographic parameters, such as tricuspid annular plane systolic excursion, RV fractional area change, and peak systolic tissue Doppler velocity, provide limited information due to the complex RV geometry. Advanced imaging techniques, including speckle-tracking, three-dimensional echocardiography, and cardiac magnetic resonance imaging, enable more accurate quantification of RV volumes, ejection fraction, strain, RV motion components and RV-pulmonary artery coupling, and have demonstrated superior prognostic value. Therefore, a comprehensive assessment of RV function using advanced imaging techniques should be incorporated into routine clinical practice to improve risk stratification and preprocedural planning before cardiac surgery and transcatheter interventions. However, it is necessary to standardize imaging protocols and define validated reference thresholds to support the clinical implementation of these state-of-the-art parameters.

Background:Right ventricular (RV) function is an important predictor of morbidity and mortality in various cardiovascular conditions. Nevertheless, its echocardiographic assessment is challenging due to its complex anatomy and location in the chest, resulting in limited inter-observer reproducibility. Objectives:We aimed to develop a novel deep learning model - EchoNet-RV - to segment the RV in apical 4-chamber view (A4C) echocardiographic videos and estimate RV fractional area change (RVFAC). Methods:For training EchoNet-RV, 7,169 expert-annotated A4C echocardiographic videos were used. The model's performance was evaluated on a held-out internal test set of 1,320 A4C videos and two international external test sets of 3,107 and 1,077 A4C videos from two separate centers. Additionally, the associations between the predicted RVFAC values and the composite endpoint of heart failure hospitalization or all-cause death were also analyzed in the first external test set. Results:EchoNet-RV segmented the RV with Dice coefficients of 0.893 (0.891-0.895), 0.797 (0.796-0.798), and 0.788 (0.785-0.790) and predicted RVFAC with mean absolute errors of 5.795 (5.560-6.031), 5.830 (5.692-5.970), and 6.362 (6.064-6.660) percentage points in the held-out test set and the two external test sets, respectively. In 500 randomly selected videos from the external test sets, EchoNet-RV's prediction error was significantly lower than the inter-observer variability (p<0.001). Moreover, it identified RVFAC <35% with areas under the receiver operating characteristic curve of 0.859 (0.843-0.876), 0.725 (0.710-0.740), and 0.684 (0.653-0.713) in the three test sets. EchoNet-RV also outperformed two multi-task models, EchoPrime and PanEcho, in estimating RVFAC and identifying RV dysfunction in the external test sets. In the first external test set, predicted RVFAC values were inversely associated with the composite endpoint (adjusted HR: 0.948 [0.917-0.979], p<0.001), independent of age, sex, cardiovascular risk factors, and left ventricular systolic function. Conclusions:EchoNet-RV enables the rapid and automated assessment of RVFAC, with strong potential to become a valuable tool for the echocardiographic evaluation of RV function and disease surveillance.

AIMS:While pre-defined reference shapes have been used to assess morphological changes in the left ventricle, standardized methods for evaluating right ventricular (RV) remodelling are lacking. This study aimed to develop and test a new 3D echocardiography (3DE)-based method for quantifying RV shape in a large cohort of healthy individuals and across various disease states. METHODS AND RESULTS:3DE-derived RV mesh models were reconstructed in 1043 healthy subjects from the World Alliance of Societies of Echocardiography (WASE) study and in 581 patients with severe aortic stenosis, heart failure with reduced ejection fraction (HFrEF), post-heart transplantation, severe primary mitral regurgitation (MR), atrial secondary tricuspid regurgitation (A-STR), tetralogy of Fallot (TOF), and pulmonary hypertension (PH). To assess global RV shape, hemi-sphericity volume ratio (HSVR) and hemi-conicity angle (HCA) were calculated, where a higher HSVR and a more acute HCA reflect more spherical and conical shapes, respectively. In the WASE population, females had more spherical RVs, whereas males had more conical RVs (P = 0.028). Considering age, younger females had more conical RVs, while older individuals in both sexes showed spherical remodelling (P < 0.05). Comparing disease groups with WASE controls, MR, HFrEF, and A-STR patients had more spherical RVs compared with controls (both P < 0.001), while PH and TOF patients showed conical remodelling (both P < 0.001). In A-STR, a more conical remodelling was associated with adverse clinical outcomes. CONCLUSION:The proposed 3DE-based method comprehensively characterizes RV geometry, demonstrating demographic variation in healthy individuals and disease-specific alterations in patients, with important prognostic implications.

AIMS:The 2025 American Society of Echocardiography guidelines on the right heart introduce severity grading for right ventricular (RV) dysfunction based on individual functional parameters. However, single-parameter assessment may result in inconsistent diagnosis of RV dysfunction and discordant grading of its severity. We aimed to investigate the prognostic value and discordance among RV functional echocardiographic parameters. METHODS:We analyzed two- and three-dimensional echocardiographic data from 3 centers, including 1,146 consecutive patients followed for the composite end point of all-cause mortality and heart failure hospitalization. RV dysfunction severity was graded using the guideline-recommended cutoff values for tricuspid annular plane systolic excursion (TAPSE), fractional area change (FAC) and free-wall longitudinal strain (FWLS) and compared with severity assessment based on RV ejection fraction (RVEF). RESULTS:Over a median follow-up of 3.2 years, 261 patients (23%) met the composite end point. Assessed by RVEF, worsening RV dysfunction categories carried a higher risk of the composite end point, which was significant between mild versus normal and moderate versus mild dysfunction (hazard ratio [HR] = 2.385 [95% CI, 1.720-3.307], P < .001; and HR = 1.581 [95% CI, 1.066-2.346], P = .023), but not between severe versus moderate dysfunction. TAPSE and FWLS did not show a significant difference in risk between the dysfunction categories. FAC identified a significant risk difference between adjacent severity categories only for moderate versus mild dysfunction (HR = 1.928 [95% CI, 1.309-2.840], P < .001). Agreement with RVEF in patients with dysfunction was poor for TAPSE (quadratic weighted κ = 0.06, P = .132) and fair for FAC and FWLS (κ = 0.34, P < .001, and κ = 0.30, P < .001, respectively). CONCLUSION:Significant discordance exists between conventional echocardiographic parameters of RV function and RVEF in grading RV systolic dysfunction. None of the individual RV functional parameters provided consistent risk stratification across all severity categories. RVEF showed the most consistent overall risk discrimination, whereas FAC and FWLS provided partial prognostic stratification.

Abstract In patients with aortic stenosis (AS) evaluating left ventricular (LV) systolic function is challenging due to the influence of increased afterload on traditional measures. Myocardial work (MW) analysis, a novel echocardiographic method, adjusts myocardial deformation to instantaneous LV pressure, providing a more accurate reflection of LV contractile state. Notably, prolonged LV pressure overload induces significant backward effects beyond the LV; the classification of this extravalvular cardiac damage effectively represents the cardiopulmonary system's involvement in AS. Both MW analysis and cardiac damage staging may possess significant prognostic value in the clinically complex cohort of transcatheter aortic valve replacement (TAVR) candidates. Thus, our objective was to evaluate the prognostic value of MW analysis and cardiac damage staging in TAVR patients. We enrolled 314 patients (79±6 years, 40% female) prior to TAVR. Echocardiographic assessments were conducted one day before the procedure. LV ejection fraction (EF) was calculated, global longitudinal strain (GLS) was measured using speckle-tracking echocardiography. LV pressure was estimated from systolic blood pressure and transaortic mean gradient, and global constructive work (GCW) was quantified using dedicated software. Based on echocardiographic data, we determined the extent of cardiac damage associated with AS, categorizing patients into Stage 0 (no cardiac damage), Stage 1 (LV damage), Stage 2 (mitral valve or left atrial damage), Stage 3 (pulmonary artery vasculature or tricuspid valve damage), or Stage 4 (right ventricular damage). The primary endpoint was all-cause mortality, reached by 69 patients during a median follow-up period of 25 months. Preprocedural EF was 47±13 %, GLS was -12.3±4.2 %, GCW was 2043±769 mmHg%. 14 (5%) patients were classified as Stage 0, 61 (20%) as Stage 1, 133 (43%) as Stage 2, 22 (7%) as Stage 3, and 74 (24%) as Stage 4. GCW showed a decline through AS Stages (from Stage 0-4: 2963±652 vs. 2154±621 vs. 2174±706 vs. 2044±827 vs. 1553±757 mmHg%; p<0.001). Using univariate Cox analysis GCW (HR 0.968 [95% CI 0.938-0.998] per 100 unit change; p=0.034) and AS Staging (HR 1.236 [95% CI 1.016-1.505]; p=0.034) were associated with all-cause mortality, while EF (HR 0.982 [95% CI 0.964-1.001]; p=NS) and GLS (HR 1.047 [95% CI 0.989-1.108]; p=NS) were not. In multivariate Cox regression models, both GCW (HR 0.958 [95% CI 0.923-0.994] per 100 unit change; p=0.022) and AS cardiac damage staging (HR 1.281 [95% CI 1.040-1.577]; p=0.020) were significant independent predictors of all-cause mortality. In TAVR patients, preoperative GCW values continuously decreased across all AS Stages. GCW and AS Staging showed strong association with all-cause mortality in our cohort, while EF and GLS did not. Furthermore, GCW and AS Staging had higher prognostic value than any other echocardiographic measure, highlighing their role in preprocedural assessment before TAVR.

Abstract Background 2D echocardiography requires multiple views for the assessment of left ventricular (LV) systolic function and does not enable the calculation of right ventricular (RV) ejection fraction (EF). 3D echocardiography (3DE) has clear incremental value over 2D echocardiography; nevertheless, its availability and feasibility is limited. Purpose We aimed to develop a dual-task deep learning model – EF2Net – for predicting 3DE-derived LVEF and RVEF from 2D apical 4-chamber (A4C) view echocardiographic videos. Methods The EF2Net model comprises two video transformers, which were first pre-trained on 29,424 unlabeled A4C videos from 15,533 echocardiographic studies in a self-supervised fashion. In the subsequent supervised training phase, one of the transformers was trained for predicting LVEF on the publicly available EchoNet-Dynamic dataset and a dual-center international 3DE dataset comprising 5,341 labeled A4C videos from 1,408 echocardiographic studies, whereas the other transformer was trained for predicting RVEF only on the latter. Beyond testing the model internally on 20% of the dual-center dataset (i.e., internal test set), it was also validated in a labeled external validation set comprising (1) 244 A4C videos of 244 patients with different cardiac diseases and available outcome data and (2) 4,421 A4C videos of 853 healthy adults from the World Alliance of Societies of Echocardiography (WASE) study. Last, we evaluated the model on a low-risk, community-based cohort (1,166 unlabeled A4C videos of 1,166 individuals) with a 10-year follow-up to investigate the associations between the predictions and all-cause mortality. Results In the internal test set and the labeled external validation set, EF2Net predicted 3DE-derived LVEF with a mean absolute error (MAE) of 4.58 and 4.67 percentage points, respectively, whereas it achieved an MAE of 4.82 and 5.43 percentage points in predicting 3DE-derived RVEF. In the labeled external validation set, the model identified an LVEF <50% and an RVEF <45% with an area under the receiver operating characteristic curve of 0.95 and 0.86, respectively. In patients with cardiac diseases, the EF2Net-predicted LVEF and RVEF values were associated with the composite of all-cause death and heart failure hospitalization (LVEF – HR: 0.94 [0.91-0.98], p=0.001; RVEF – HR: 0.92 [0.88-0.98], p=0.004) independent of age and sex. Moreover, in the community-based cohort, the EF2Net-predicted EF values were also associated with 10-year all-cause mortality (LVEF – HR: 0.97 [0.95-0.99], p=0.026; RVEF – HR: 0.91 [0.88-0.95], p<0.001) independent of age, sex, and LV diastolic function. Conclusions EF2Net enabled the automated and accurate assessment of biventricular systolic function based on a single 2D echocardiographic view. It also exhibited robust performance when validated in a multi-ethnic dataset, and the prognostic value of its predictions was confirmed in patients with cardiac diseases and in the community.

Background: Orthotopic heart transplantation (OHT) remains the gold standard for end-stage heart failure, yet individualized risk assessment for postoperative mortality remains challenging. We aimed to develop and interpret random forest-based models for predicting 30-day and 1-year mortality and to examine whether the key predictors differ between the 30-day and 1-year models. Methods: We analyzed 581 patients who underwent OHT between 2012 and 2024. The 30-day and 1-year mortality rates were 9.9% and 17.6%, respectively. Eighty-seven preoperative and forty-eight postoperative variables were considered as input features for model development. Random forest models were trained and validated using five-fold cross-validation, and explainability was assessed using SHapley Additive exPlanations (SHAP). Results: Using preoperative features only, the random forest models achieved AUCs of 0.62 (95% CI, 0.48–0.75) for 30-day and 0.67 (95% CI, 0.56–0.78) for 1-year mortality. SHAP analysis revealed that early mortality predictions were primarily driven by features reflecting acute physiological stress—hepatic dysfunction, inflammation, and hemodynamic instability—whereas long-term predictions were increasingly influenced by renal function, metabolic reserve, and frailty. Incorporating postoperative features improved performance (AUC 0.98 [95% CI, 0.97–0.99] and 0.86 [95% CI, 0.80–0.92], respectively), with model predictions dominated by the severity and persistence of organ dysfunction: short-term risk driven by hepatic injury, hemodynamic compromise, and critical illness, and long-term risk by sustained hepatic and renal impairment, metabolic resilience, and duration of circulatory support. Conclusions: Random forest models integrating preoperative and immediate postoperative data could predict short- and mid-term mortality after OHT. SHAP analysis demonstrated temporal shifts in the most important predictors, supporting the role of dynamic, data-driven risk assessment in transplant care.

BACKGROUND: Right ventricular (RV) function has a well-established prognostic role in patients with severe mitral regurgitation (MR) undergoing transcatheter edge-to-edge repair (TEER) and is typically assessed using echocardiography-measured tricuspid annular plane systolic excursion. Recently, a deep learning model has been proposed that accurately predicts RV ejection fraction (RVEF) from 2-dimensional echocardiographic videos, with similar diagnostic accuracy as 3-dimensional imaging. This study aimed to evaluate the prognostic value of the deep learning–predicted RVEF values in patients with severe MR undergoing TEER. METHODS: This multicenter registry study analyzed the associations between the predicted RVEF values and 1-year mortality in patients with severe MR undergoing TEER. To predict RVEF, 2-dimensional apical 4-chamber view videos from preprocedural transthoracic echocardiographic studies were exported and processed by a rigorously validated deep learning model. RESULTS: Good-quality 2-dimensional apical 4-chamber view videos could be retrieved for 1154 patients undergoing TEER between 2017 and 2023. Survival at 1 year after TEER was 84.7%. The predicted RVEF values ranged from 26.6% to 64.0% and correlated only modestly with tricuspid annular plane systolic excursion (Pearson R =0.33; P <0.001). Importantly, predicted RVEF was superior to tricuspid annular plane systolic excursion levels in predicting 1-year mortality after TEER (area under the curve, 0.687 versus 0.625; P =0.029). Furthermore, Kaplan-Meier survival analysis revealed that patients with reduced RV function (n=723; defined as a predicted RVEF of <45%) had significantly worse 1-year survival rates than patients with preserved RV function (n=431; defined as a predicted RVEF of ≥45%; 80.3% [95% CI, 77.4%–83.3%] versus 92.1% [95% CI, 89.5%–94.7%]; hazard ratio for 1-year mortality, 2.67 [95% CI, 1.82–3.90]; P <0.001). CONCLUSIONS: Deep learning–enabled assessment of RV function using standard 2-dimensional echocardiographic videos can refine the prognostication of patients with severe MR undergoing TEER. Thus, it can be used to screen for patients with RV dysfunction who might benefit from intensified follow-up care.

Abstract Background/Introduction Accurate assessment of left ventricular (LV) and right ventricular (RV) systolic function is crucial in managing people with congenital heart disease (CHD), but can be challenging given the vast anatomic heterogeneity and varied surgical history in this population. While 3D echocardiography (3DE) is the most reproducible and best-validated echocardiographic technique for assessing biventricular systolic function, it requires special equipment and expertise and is not feasible in patients with poor acoustic windows. EF2Net, a dual-task deep learning model, has shown promise in predicting 3DE-derived LV and RV ejection fraction (EF) from standard 2D echocardiographic (2DE) 4-chamber views (1). The model has been previously tested in patients with acquired heart disease, healthy volunteers, and a low-risk community-based cohort but has not been validated in adults with CHD yet. Purpose We sought to validate the performance of the EF2Net model in predicting LV and RV EF in adults with CHD. Methods Ninety-six consecutive adults with CHD who had undergone echocardiography as part of their routine clinical follow-up were screened. Patients with univentricular physiology and systemic RV were excluded, as were those with insufficient image quality to obtain 3DE full-volume datasets for both ventricles. The final cohort comprised 90 patients (177 apical 4-chamber views). The EF2Net deep learning model was applied and its predictions of LV and RV EF were compared to the actual 3DE measurements (Figure). Results The median age of the cohort was 28.0[IQR 23.0-34.5] years and 48.9% were female. The most common CHD diagnosis was a shunt lesion in 37.8% (of which 64.7% included a post-tricuspid component), followed by LV outflow tract disease in 28.8%, and RV outflow tract disease in 24.4%. Most patients (90%) were in NYHA functional class I. The mean 3DE LV EF was 60.9±6.8% and the mean 3DE RV EF was 51.5±6.8%. EF2Net predicted 3DE-derived LV EF and RV EF with a mean absolute error (MAE) of 5.0 and 6.4 percentage points, respectively. Conclusions The EF2Net model accurately predicts biventricular EF in adults with CHD using routine 2DE 4-chamber views. This validation underscores the potential of EF2Net to enhance clinical practice by providing reliable, non-invasive assessments of ventricular function in patients with complex ventricular morphology, particularly in settings where 3DE is not readily available.

Left ventricular noncompaction (LVNC) is characterized by excessive trabeculation, which may impair left ventricular function over time. While cardiac magnetic resonance imaging (CMR) is considered the gold standard for evaluating LV morphology, the optimal modality for follow-up remains uncertain. This study aimed to assess the correlation and agreement among two-dimensional transthoracic echocardiography (2D_TTE), three-dimensional transthoracic echocardiography (3D_TTE), and CMR by comparing volumetric and strain parameters in LVNC patients and healthy individuals. Thirty-eight LVNC subjects with preserved ejection fraction and thirty-four healthy controls underwent all three imaging modalities. Indexed end-diastolic, end-systolic, and stroke volumes, ejection fraction, and global longitudinal and circumferential strains were evaluated using Pearson correlation and Bland–Altman analysis. In the healthy group, volumetric parameters showed strong correlation and good agreement across modalities, particularly between 3D_TTE and CMR. In contrast, agreement in the LVNC group was moderate, with lower correlation and higher percentage errors, especially for strain parameters. Functional data exhibited weak or no correlation, regardless of group. These findings suggest that while echocardiography may be suitable for volumetric follow-up in LVNC after baseline CMR, deformation parameters are not interchangeable between modalities, likely due to trabecular interference. Further studies are warranted to validate modality-specific strain assessment in hypertrabeculated hearts.

In routine clinical practice, a vast amount of data is generated, including myriads of ultrasound recordings. However, their annotation and interpretation are labor-intensive; thus, a method that can incorporate this unlabeled data into deep learning pipelines would be highly beneficial. Video masked autoencoders (VideoMAE) are state-of-the-art pre-training techniques and have performed exceptionally well in various computer vision tasks. Accordingly, we hypothesized that a VideoMAE pre-trained on a large unlabeled dataset of ultrasound recordings could also perform well in a downstream task following supervised training on a smaller but labeled dataset. Nevertheless, we found that the conventional masking strategy of the VideoMAE pipeline may perform sub-optimally in the specific domain of ultrasound videos. Motivated by this, we proposed a novel region of interest (ROI)-aware masking method that considers the specific characteristics of this domain. We demonstrated that applying our method instead of the conventional masking strategy significantly improves the VideoMAE's performance in clinically relevant downstream tasks, even when we reduced the labeled training dataset to one-tenth of its original sample size. The source code for this paper is available at https://github.com/szadam96/ROI-aware-masking.