Deep learning enables fully automated cineCT-based assessment of regional right ventricular function

Abstract

Aims

Right ventricular (RV) function is a key factor in the diagnosis and prognosis of heart disease. However, current advanced computed tomography (CT)-based assessments rely on semi-automated segmentation of the RV blood pool and manual delineation of the RV free and septal wall boundaries. These steps are time-consuming and prone to inter- and intra-observer variability.

Methods and results

We developed and evaluated a fully automated pipeline consisting of two deep learning methods to automate volumetric and regional strain analysis of the RV from contrast-enhanced, electrocardiogram (ECG)-gated cineCT images. The Right Heart Blood Segmenter (RHBS) is a 3D high-resolution configuration of nnU-Net to define the endocardial boundary, while the Right Ventricular Wall Labeler (RVWL) is a 3D point cloud-based deep learning method to label the free and septal walls. We trained our models using a diverse cohort of patients with different RV phenotypes and tested them in an independent cohort of patients with aortic stenosis undergoing TAVR. Our approach demonstrated high accuracy in both cross-validation and independent validation cohorts. RHBS and RVWL both yielded Dice scores of 0.96 and accurate volumetry metrics. RVWL achieved high Dice scores (>0.90) and high accuracy (>93%) for wall labelling. The combination of RHBS + RVWL provided an accurate assessment of free and septal wall regional strain, with a median cosine similarity value of 0.97 in the independent cohort.

Conclusion

A fully automated 3D cineCT-based RV regional strain analysis pipeline has the potential to significantly enhance the efficiency and reproducibility of RV function assessment, enabling the evaluation of large cohorts and multi-centre studies.

Graphical Abstract

Graphical Abstract.

Key points.

1. RV endocardial segmentation of contrast-enhanced CT scans can be utilized to perform volumetry, and when paired with labelling of free and septal walls, regional evaluation of surface strain.

2. However, this has previously been performed using time-intensive semi-automated segmentation methods and manually labelling free wall and septal wall regions.

3. Here, we describe an automated, deep learning-based approach that uses two separate DL models to define the endocardial boundary (in 3D) and then label the free and septal walls on the endocardial surface.

4. Our approach facilitates rapid and automatic advanced phenotyping of patients. This reduces prior limitations of potential interobserver variability and challenges associated with evaluating large cohorts.

Introduction

Right ventricular (RV) volumetry and systolic performance are essential metrics for diagnosis and prognosis of both right- and left-sided heart disease. For example, RV volumetry informs interventional decisions for valvular disease,^1–3 and helps guide long-term management of congenital heart disease like tetralogy of Fallot.^2,4 Pre-operative RV dysfunction in the setting of left ventricular (LV) failure also predicts RV failure after left ventricular assist device (LVAD) implantation.^5,6 In addition to volumetry metrics, our group has used ECG-gated cineCT to estimate regional RV strain in patients with various pathophysiologies of RV dysfunction.^7–9 These assessments highlight patient-specific preload and afterload features that contribute to RV function.

However, current assessments rely on semi-automated segmentation of the RV blood pool and manual delineation of the RV free and septal wall boundaries.^7–9 This manual process limits clinical translation due to inter- and intra-observer variability, reduced scalability, and challenges in multi-centre study. Several technical factors impair automated analysis, including the complex, crescent-like shape of the RV, the high prevalence of metal artefacts due to replacement valves or pacing wires, and the need for high spatial resolution segmentation for downstream regional evaluation.

Recently, deep learning (DL) algorithms have enabled accurate and reproducible 3D cardiac image segmentation. However, methods for RV segmentation tailored for regional analysis are lacking. Existing algorithms often focus on left heart structures,^10,11 exclude diverse disease phenotypes,^12–14 or analysed cardiac magnetic resonance (MR) images.^15,16 Some RV segmentation methods use 2D slice-by-slice approaches, smooth 3D boundary predictions, or significantly downsample imaging volume.^12,15,16 Here, we utilize nnU-Net V2,¹⁷ a robust semantic segmentation approach,¹⁸ to segment the RV and evaluate the accuracy of our Right Heart Blood Segmenter (RHBS) using volumetric measures as well as the ability to provide comparable strain metrics.^8,19,20

To interpret regional strain differences within the RV, delineation of the free and septal walls is needed.^8,9 Previously, these surfaces were manually delineated during the end-diastole timeframe. While DL architectures exist for labelling 3D point clouds,^21,22 they have not been applied to the RV. We present the Right Ventricular Wall Labeler (RVWL), a DL model for automated labelling of the RV free and septal walls, and evaluate its accuracy and impact on strain analysis.

In this work, we develop and evaluate a fully automated pipeline using two DL models to perform volumetric and regional strain analysis of the RV from contrast-enhanced, ECG-gated cineCT images. We address three unanswered questions: 1) Can RHBS, a 3D convolutional neural network (CNN) trained on a heterogeneous patient cohort, accurately segment the RV blood pool for volumetry and regional strain analysis? 2) Can RVWL, a 3D point cloud-based DL model, accurately label the RV free and septal walls? 3) Does the combined RHBS + RVWL pipeline provide accurate regional strain assessments? We validated our models using cross-fold validation and tested them in an independent cohort of patients undergoing TAVR to assess generalizability.

Methods

The overall data processing pipeline is shown in Figure 1. The utility of such an automated pipeline is shown in Figure 2.

Figure 1.

Overview of processing pipeline. Right Heart Bloodpool Segmenter (RHBS) is a nnU-Net V2 based deep learning (DL) model trained to automatically segment the right ventricle (RV), right artium (RA), pulmonic valve (PV), and pulmonary artery (PA) bloodpools from ECG-gated cineCT images. Then, RV Wall Labeller (RVWL) delineates the free and septal walls of the RV bloodpool boundary.

Figure 2.

Example of quantitative evaluation of right ventricular (RV) free and septal wall strain. Endocardial surface tracking from RV segmentation yields a 3D map of regional strain (left). Each patch on the surface is then categorized into one of 6 groups based on the end-systolic strain value to obtain a profile of strain per patient (histogram and map below). Strain profiles for the free wall and septal wall are obtained by reporting the strains from the patches included in the predicted free wall and septal wall labels (table on right).

Training population

With IRB approval, 101 ECG-gated, contrast-enhanced cardiac CT scans obtained between 07/2017 and 11/2023, which were manually processed as part of earlier studies, were used for model training.^5,6,8,9 Cases used in training were identified from several clinical populations who routinely undergo cardiac cineCT at our institution: patients undergoing anthracycline therapy (AN, n = 8), pre-operative assessment of chronic thromboembolic pulmonary hypertension (CTEPH, n = 14), left-sided heart failure patients undergoing evaluation for LVAD implantation (LVHF, n = 50), and adults with congenital heart disease undergoing evaluation for transcatheter pulmonary valve replacement (ACHD, n = 29).

All patients underwent a full cycle, ECG-gated cineCT on a 256-slice Revolution CT scanner (Waukesha, WI). Gantry rotation speed was 280 ms for nearly all cases (n = 1 acquired at 234 ms). Based on the clinical imaging protocol, patients were scanned at 80, 100, or 120 kV, and the maximum tube current ranged from 137 to 781 mA. Nearly all images were reconstructed with 0.625 mm slice thickness (n = 6 reconstructed with 1.25 mm slices). Most axial images were reconstructed at ∼10% intervals (n = 23 reconstructed at <9% intervals) of the cardiac cycle (0–90% of the R–R interval). However, 6 cases covered 0 - < 85% of the cardiac cycle. Images were reconstructed on a 512 × 512 matrix using a standard reconstruction kernel. The field of view was typically 200 mm (range: 179–360 mm), which led to an in-plane pixel size of 0.39 mm × 0.39 mm. RV enhancement ranged from 203–1057 Hounsfield units. All cases had sufficient image quality for manual segmentation. Images were not excluded due to the presence of metal artifact (n = 49 had metal artifacts present in the RV).

Two frames per patient were selected for training: the end-diastolic frame and an estimation of the end-systolic frame based on the total number of frames. Therefore, the total training dataset consisted of 202 images.

Test population

Model accuracy was also evaluated in an independent cohort of patients with severe aortic stenosis undergoing evaluation for transcatheter aortic valve replacement (n = 24). Scans occurred between 12/2019 and 11/2022. After review of all scans (n = 290), 24 cases were selected to achieve a balanced distribution of cases with varying RV enhancement.

All patients underwent a full cycle, ECG-gated cineCT on a 256-slice Revolution CT scanner (Waukesha, WI). Gantry rotation speed was 280 ms. Based on the clinical imaging protocol, patients were scanned at 80, 100, or 120 kV and the maximum tube current ranged from 402–719 mA. Nearly all images were reconstructed with 0.625 mm slice thickness (n = 1 reconstructed with 1.25 mm slices). 11 cases had axial images reconstructed at ∼10% intervals of the cardiac cycle, while 13 cases were reconstructed at <9% intervals. Images were reconstructed on a 512 × 512 matrix using a standard reconstruction kernel. The field of view was typically 250 mm (range: 190–271 mm), which led to an in-plane pixel size of 0.49 mm × 0.49 mm. RV enhancement ranged from 218–1207 Hounsfield units. Seven cases had metal artifacts present in the RV.

Semi-automated expert bloodpool segmentation and metrics of RV function

Voxel-wise manual segmentation of the RV, right atrium (RA), pulmonic valve (PV), and pulmonary artery (PA) blood volumes on images at the native resolution was confirmed by a trained operator with 6 years of experience in cardiac image segmentation (author A.C.) using ITK-SNAP (Philadelphia, PA). RV volumes were obtained from each segmentation. Global RV function was measured as end-diastolic volume (EDV) and end-systolic volume (ESV), while regional RV function was measured using regional strain (RS_CT). Mean RV RS_CT and mean RS_CT in the free and septal walls were measured. RS_CT values were also categorized by the distribution of strain strength at end-systole. Methods for measuring and categorizing RV RS_CT from cardiac cineCT have been previously described in detail.^7,8

Automated delineation of the RV endocardial volume/boundary via RHBS

For uniformity, images were resampled to 1 mm isotropic voxels. RHBS labelled each voxel as one of five labels: RV, RA, PV, PA, and background. We trained the model using nnU-Net V2’s 3D full resolution configuration¹⁷ on a 16-core Ubuntu computer with 128 GB RAM and a 24 GB NVIDIA RTX A5000 (NVIDIA Corporation, Santa Clara, CA). Training and validation were performed using five-fold stratified cross-validation to ensure each fold included a representative sample of each clinical population. Each fold was trained for 1000 epochs per nnU-Net default.

Metrics of RHBS accuracy

RV segmentation accuracy was evaluated volumetrically with the Dice coefficient. Global volumetry metrics were evaluated using the absolute and absolute percent errors of RV end-diastolic and end-systolic volumes as well as the difference in mean end-systolic RV strain measured with the predicted segmentations relative to the manual approach.

Regional function accuracy was measured as the cosine similarity metric (CSM) between the end-systolic RS_CT categorized with the manual and predicted segmentations. Cosine similarity measures the similarity between the directions of two vectors in space. Given that the RS_CT categorization is reported as the percent of the RV surface that belongs to each category,⁸ this distribution was treated as a vector in multi-dimensional space. Cosine similarity evaluated the similarity of the direction of two vectors; in this case, the vectors described the RS_CT categorizations of the reference segmentation, and RHBS predicted segmentation.

Manual labelling of free and septal walls

The end-diastolic phase of the ECG-gated cineCT dataset was used for labelling of the free and septal wall surfaces.^8,9 To do so, the semi-automated segmentation of the RV blood pool was visualized using ITK-SNAP (Philadelphia, PA). Each surface was then delineated by a trained operator with 6 years of experience in cardiac image segmentation (author A.C.). The labelling does not affect volumetric or global strain analysis. It is used to select the portions of the RV endocardial mesh for analysis of regional strain.

Automated labelling of free and septal walls via RVWL

The semi-automatic RV bloodpool segmentations with free wall and septal wall labels were converted into point clouds for automated free and septal wall labelling. Each patient yielded a set of surface points that were labelled as either being the ‘free wall’, ‘septal wall’, or ‘none’.

PointNet’s architecture maps each independent point within a point cloud to a high-dimensional feature and uses maxpooling to aggregate these features into a global descriptor.²¹ PointNet++ extends this approach by introducing hierarchical set abstraction, which enables the model to learn local geometric features with spatial coherence.²² Multi-Scaling Grouping (MSG) creates several neighbourhoods that each have a different radius, allowing PointNet++ to capture more detailed information about both fine-grained local structures and broader contextual geometry of the Free Wall and Septal Wall.

RVWL is a PointNet++ with MSG network trained to predict the labels generated by human annotation. Each scan yields 1 training example, so the total training dataset was 101 point clouds. PointNet++ analyses 3000 points (from a point cloud) at a time, but our surfaces had between 11 405 and 36 789 points. Therefore, surface points were subsampled into 3000 point sets, which served as an augmentation during training. The network was trained using the PyTorch for 1000 epochs. The same cross-fold partitions defined for the 3D U-net were used for PointNet++ training and cross-fold evaluation.

Post-processing of predictions was performed to minimize mislabelling of the walls due to weak predictions. Specifically, after prediction, free wall and septal wall labels were only kept if the class label probability ≥ 0.9. Otherwise, these weak predictions were relabelled as ‘none’.

Wall labelling performance was evaluated with the Dice score of each label (free wall, septal wall, none). Additionally, overall accuracy and extent of mislabelling (i.e. a true septal wall point predicted as free wall) were also evaluated. Accuracy of the resulting regional function was measured as the correlation of the mean end-systolic strain of the free wall and septal wall with respect to the standard approach. Regional function accuracy was also measured as the cosine similarity between the predicted strain profiles for each wall and the standard approach.

We compared the performance of our pipeline to intra- and inter-observer variability in global and regional RV function, as described in C.

Statistics

Normality of datasets was evaluated with the Shapiro-Wilk test. For nonparametric datasets, statistical differences in prediction results between the cross-validation cohort and the independent cohort were evaluated with a two-sided Wilcoxon rank sum test. Differences between populations were computed with the Kruskal–Wallis test for nonparametric datasets. Correlation studies between nonparametric datasets were computed with the Spearman correlation. Pearson correlation was used to evaluate normally distributed data. Statistical testing was computed with MATLAB (MathWorks, Natick, MA).

Results

Patient demographics

Patient demographics are summarized in Table 1. TAVR recipients were older (83 years, IQR: 74–91) than the AN (59 years, IQR: 40–69, P = 0.01), CTEPH (57 years, IQR: 49–61, P < 0.01), LVHF (59 years, IQR: 50–68, P < 0.01), and ACHD (32 years, IQR: 24–35, P < 0.01) cohorts. The ACHD patients were younger than the AN (P = 0.04), CTEPH (P < 0.01), and LVHF (P < 0.01) patients.

Table 1.

Demographic and distribution of patients amongst training folds

​	Cross-Fold Training and Validation​				Indep. Validation​	​	​
​	AN​ N = 8​	CTEPH​ N = 14​	LVHF​ N = 50​	ACHD​ N = 29​	TAVR​ N = 24​	Total​ N = 125​	P-value​
Demographics​
Age, years​	59​ (40–69)​	57​ (49–61)​	59​ (50–68)​	32​ (24–35)​	83​ (74–91)​	57 (40–70)​	<0.01​
Sex, n (%)​	7 (88)​	8 (57)​	6 (12)​	14 (48)​	9 (38)​	44 (35)​	<0.01​
HR (bpm)​	61​ (53–66)​	76​ (65–80)​	81​ (70–95)​	72​ (61–81)​	63​ (57–76)​	74​ (63–86)​	<0.01​
Distribution Amongst Folds​
Fold 1, n (%)​	2 (10)​	3 (14)​	10 (50)​	6 (29)​	−​	21 (21)​	−​
Fold 2, n (%)​	2 (10)​	3 (15)​	10 (50)​	5 (25)​	−​	20 (20)​	−​
Fold 3, n (%)​	2 (10)​	2 (10)​	10 (50)​	6 (30)​	−​	20 (20)​	−​
Fold 4, n (%)​	1 (5)​	3 (15)​	10 (50)​	6 (30)​	−​	20 (20)​	−​
Fold 5, n (%)​	1 (5)​	3 (15)​	10 (50)​	6 (30)​	−​	20 (20)​	−​

Transcatheter aortic valve replacement (TAVR) recipients were older than the anthracycline therapy (AN), chronic thromboembolic pulmonary hypertension (CTEPH), left ventricular heart failure (LVHF), and adult congenital heart disease (ACHD) cohorts while the ACHD cohort was younger than the AN, CTEPH, and LVHF cohorts. The LVHF cohort had fewer women than the AN, CTEPH, and ACHD cohorts, and had higher heart rates than the AN and TAVR cohorts. Five-fold cross validation distributed the populations across the five folds as evenly as possible.

The LVHF cohort had fewer women (n = 6/50) than the AN (n = 7/8, P < 0.01), CTEPH (n = 8/14, P = 0.02), and ACHD (n = 14/29, P = 0.01) cohorts.

Median heart rate was 74 beats per minute (IQR: 63–86). The LVHF cohort had higher heart rates (81 bpm, IQR: 70–95) than the TAVR (63 bpm, IQR: 57–76, P < 0.01) and AN (61 bpm, IQR: 53–66, P = 0.03) cohorts.

RHBS training and quantitative evaluation

Training:

Fold 1 was trained on 80 cases and validated on 21 cases (160 and 42 volumes, respectively). Folds 2–5 were trained on 81 cases and validated on 20 cases (162 and 40 volumes, respectively). The distribution of each population within each fold is demonstrated in Table 1. The independent validation dataset was predicted using the model from Fold 3.

Quantitative Evaluation:

RHBS performance results are shown in Table 2. Dice was high and comparable in the cross-validation (0.96, IQR: 0.95–0.97) and independent validation cohorts (0.96, IQR: 0.94–0.97, P > 0.05). Absolute volume error and percent volume error were not different between the cross validation (absolute error: 7 mL, IQR: 3–12; percent error: 4%, IQR: 2–8) and independent validation cohorts (absolute error: 5 mL, IQR: 2–9, P > 0.05; percent error: 5%, IQR: 3–7, P > 0.05). 95% of cases were segmented with Dice > 0.90 and percent volume error < 20%. End-systolic frames had lower Dice scores than end-diastolic frames for both the cross-validation (end-diastole: 0.97, IQR: 0.96–0.98; end-systole: 0.95, IQR: 0.94–0.96; P < 0.01) and independent validation (end-diastole: 0.97, IQR: 0.96–0.97; end-systole: 0.94, IQR: 0.93–0.96; P < 0.01) cohorts (Figure 3). We also evaluated the impact of RV enhancement and clinical population on RHBS accuracy (Figure 3). While we did not find a significant effect with RV enhancement, there was a statistically significant difference in Dice scores across populations (P = 0.02), with a trend of higher scores in AN and CTEPH patients and lower scores in ACHD and TAVR subjects. However, these differences were not significant on post-hoc comparison.

Table 2.

Right Heart Blood Segmenter (RHBS) cross-validation and independent validation result metrics

Metric​	Cross Validation​ (n = 202)​	Independent Validation (n = 48)​	P-value​​
Dice Score​	0.96 (0.95–0.97)​	0.96 (0.94–0.97)​	0.06
Absolute Volume Error (mL)​	7 (3–12)​	5 (2–9)​	0.07
% Volume error​	4 (2–8)​	5 (3–7)​	0.08​
Cosine similarity​	0.98 (0.95–0.99)​	0.99 (0.95–0.99)​	0.10​

Dice scores, absolute volume error, percent volume error, and cosine similarity were similar in the cross validation and independent validation cohorts.

Figure 3.

Dice score evaluated by cardiac phase, level of right ventricular (RV) enhancement, and clinical population. End-diastolic frames had higher Dice scores than the end-systolic frames in both the cross validation and independent validation cohorts. Dice scores were not different across quintiles of RV enhancement in the cross validation or independent validation cohorts. Dice scores were also not significantly different between populations. Grey dashed line indicates the 25th percentile of Dice scores for each dataset.

RHBS yielded mean RV strain measures which agreed with values derived from semi-automated, expert segmentations (Figure 4, left). Pearson correlation was 0.97 for the cross-validation cohort and 0.98 for the independent validation cohort, and they were not significantly different from one another (P = 0.54). The root mean squared error was 0.02 and 0.01 for the cross-validation and independent validation cohorts, respectively.

Figure 4.

Accuracy of DL-based Right Heart Bloodpool Segmenter (RHBS)-derived mean RV strain (left) and strain categorization (right). Left: Deep learning (DL)-derived segmentation agreed very strongly in both the cross-validation (blue) and independent validation (red) cohorts with semi-automated right ventricular (RV) segmentation-based analysis. Right: In addition to overall mean strain, categorization of the RV strain very closely (>0.98) agreed with semi-automated analysis and was comparable between both validation cohorts.

RHBS yielded an accurate categorization of RS_CT. Cosine similarity (Figure 4, right) was similar between both testing cohorts (cross-fold validation: 0.98, IQR: 0.95–0.99; independent: 0.99, IQR: 0.97–0.99, P > 0.05).

RVWL training and quantitative evaluation

Training:

The best performing RVWL for each fold had a validation dice of 0.91 or 0.92, and occurred between epoch 322 and 889. The independent validation dataset was predicted using the model from fold 3.

Quantitative Evaluation:

Table 3 summarizes the quantitative evaluation of RVWL. Automated wall labelling via RVWL yielded similar distributions of the RV surface labelled FW, SW, or None compared to semi-automated labelling by an expert user. By semi-automated RV wall labelling, the median percentage of the RV wall surface labelled FW was 43% (IQR: 41–46), 16% (IQR: 15–18) was labelled SW, and 41% (IQR: 37–43) was labelled None. RVWL resulted in more of the RV labelled SW 20% (IQR: 18–21, P < 0.01) and less of the RV labelled None 36% (IQR: 34–39, P < 0.01) compared to the reference. RVWL-derived FW labelling was comparable to the reference.

Table 3.

Right Ventricle Wall Labeller (RVWL) cross-validation and independent validation result metrics

Metric​	​	Cross validation​ (n = 101)​	Independent validation​ (n = 24)​	P-value​
RVWL Dice​	FW​	0.94 (0.92–0.95)​	0.93 (0.91–0.94)​	0.13​
	SW​	0.92 (0.90–0.94)​	0.93 (0.92–0.94)​	0.68​
	None​	0.90 (0.88–0.92)​	0.91 (0.89–0.92)​	0.67​
RVWL Mesh Accuracy​	FW​	95 (93–97)​	96 (93–98)​	0.31​
	SW​	94 (91–97)​	93 (89–95)​	0.06​
RVWL % Mislabeled​	FW​	0 (0 - < 1)​	0 (0 - < 1)​	0.48​
	SW​	<1 (0–1)​	1 (<1–2)​	<0.01​
RVWL CSM​	FW​	0.99 (0.99–0.99)​	0.99 (0.99–0.99)​	0.12​
	SW​	0.99 (0.99–0.99)​	0.99 (0.99–0.99)​	0.06​
RHBS  +  RVWL CSM​	FW​	0.97 (0.93–0.99)​	0.97 (0.94–0.99)​	0.77​
	SW​	0.97 (0.94–0.99)​	0.97 (0.93–0.98)​	0.20

Dice and accuracy of point labelling was high across walls and validation groups. Percentage of points mislabelled (either true septal wall, SW, point labelled free wall, FW, or vice versa) was low although statistically higher in the independent validation for SW points. Cosine similarity metric (CSM) of strain analysis was high across walls and cohorts for both RVWL based labelling and the combination of right heart blood segmenter (RHBS) segmentation followed by RVWL labelling.

Dice and mesh accuracy for all labels across both cross-validation and independent testing cohorts was high (>0.90 and >93%, respectively). Mislabelling (SW points labelled FW and vice-versa) on the RV mesh was low (≤1%), however, independent validation had a small, but statistically greater number of septal wall points mislabelled as free wall (1%, IQR: <1–2) compared to the cross-validation dataset (<1%, IQR: 0–1, P < 0.01). RVWL applied to manual segmentations resulted in high CSM (0.99 in both the cross-fold and independent testing cohorts).

Dice scores were different across labels (FW: P = 0.03, SW: P < 0.01, None: P < 0.01) for all populations. SW Dice was greater for AN (0.93, IQR: 0.92–0.94, P < 0.05) and LVHF (0.93, IQR: 0.92–0.95, P < 0.01) patients than ACHD (0.91, IQR: 0.88–0.92). SW Dice was also greater for LVHF than CTEPH (0.90, IQR: 0.89–0.93, P = 0.01). None label Dice was greater in LVHF (0.91, IQR: 0.89–0.93) than ACHD (0.89, IQR: 0.86–0.91, P < 0.01) patients. However, FW Dice was not different between populations on post-hoc analysis.

Subsequently, the combination of RHBS and RVWL was evaluated. Mean RV strain derived from RHBS + RVWL was very strongly correlated to mean strain derived from the semi-automated approach (Figure 5, top). Mean strain Pearson correlations were not different between the cross validation or independent validation cohorts in the free wall (Cross: r = 0.96, P < 0.01; Indep: r = 0.95, P < 0.01; r-to-z transformation P = 0.85) or septal wall (Cross: r = 0.96, P < 0.01; Indep: r = 0.93, P < 0.01; r-to-z transformation P = 0.78). CSM for the combined approach of RHBS + RVWL is reported in Table 3. Median CSM was 0.97 in the independent validation. The accuracy of the full automated approach is further illustrated in Figure 5, bottom. See Supplementary data online for assessment of inter- and intra-observer reproducibility.

Figure 5.

Accuracy of RHBS  +  RVWL-derived FW and SW strain in both cross-fold validation and independent testing cohort. Top: The combination of both deep learning (DL)-based right heart blood segmenter (RHBS) and DL-based right ventricule wall labeller (RVWL) provides mean free wall (FW) and septal wall (SW) strain estimates which closely match standard semi-automated processing. Bottom: Classification of surface strain into different categories with RHBS + RVWL closely matches the reference method.

Representative Examples Automated RV volumetry and strain mapping for two representative patients are shown in Figure 6 (left: cross-validation example, right: independent validation cohort example).

Figure 6.

Accuracy of quantitative RV analysis with RHBS  +  RVWL-derived metrics. Left: Results from a representative cross validation patient. Volumetry errors due to right heart blood segmenter + right ventricle wall labeller (RHBS + RVWL)-based analysis in this left ventricular heart failure (LVHF) case were 3 mL for RV end-diastolic volume (RVEDV), 5 mL for RV end-systolic volume (RVESV), 2 mL for RV stroke volume (RVSV), and 1% for RV ejection fraction (RVEF). Mean strain in both the free wall was the same but slightly lower in the septal wall compared to the expert annotation. The distribution of strain categories is similar to reference values (histograms and RV maps). Cosine similarity is 0.98 for the free wall and 0.99 for the septal wall. Right: Results from a representative independent validation (transcatheter aortic valve replacement, TAVR, cohort) case. Volumetry errors were 8 mL for RVEDV, 5 mL for RVESV, 3 mL for RVSV, and 1% for RVEF. Regional mean strain in the free wall and septal wall was slightly greater than the reference. The distribution of strain categories is similar to reference (histograms). Cosine similarity is 0.96 for the free wall and 0.87 for the septal wall. 3D strain maps demonstrate similar distributions.

The cross-validation case is an LVHF patient. Absolute differences between the manual and RHBS-derived segmentation were 3 mL for RVEDV, 5 mL for RVESV, 2 mL for RVSV, and 1% for RVEF. The distribution of the strain mapping histograms is similar, with both maps showing the percent of the RV surface in each category in descending order as Akinetic, Hypokinetic, Low Kinetic, Dyskinetic, High Kinetic, and Hypokinetic. 3D strain mapping shows a similar spatial distribution of RV strain, with large areas of Akinetic strain surrounding patches of Dyskinetic strain in the free and septal walls.

In the independent validation example, absolute differences between the manual and RHBS-derived segmentation were 8 mL for RVEDV, 5 mL for RVESV, 3 mL for RVSV, and 1% in RVEF. The distribution of RV strain matches in both maps. 3D strain maps were also similar; however, the map derived from RHBS + RVWL had more hyperkinetic areas in the free wall and hypokinetic areas in the septal wall than the reference method.

Discussion

In this study, we describe how two neural networks, RHBS and RVWL, trained on a diverse set of RV phenotypes, can be combined to automate regional RV strain assessment. RHBS enables automated RV bloodpool segmentation while RVWL labels free and septal wall points on the RV surface. After cross-validation, we tested the pipeline in an independent cohort of patients undergoing TAVR. In both cohorts, the combined pipeline yielded volumetry, mean RV FW and SW strains, and strain distributions that closely matched the standard processing pipeline. This represents a significant advancement as automated analysis has been limited by the lack of high-throughput methods and the need for manual processing.

The reproducibility of RV volumetry on cardiac CT has historically been limited by lower temporal resolution and the challenge of delineating trabeculations compared to MRI. However, our automated pipeline achieved a median Dice score of 0.96, which exceeds expert inter-observer variability of RV volumetric measurements with CT²³ and MRI²⁴ (Dices scores ranging from 0.89 to 0.92 and 0.90 to 0.93, respectively).

Clinically, the precision of our RV functional assessment is robust. Given that normal RV EDV on CT can exceed 150 mL,²⁵ our median absolute error of 5 mL represents a deviation of only ∼3%, which is negligible for clinical decision-making. This error margin is superior to the reported limits of agreement between manual CT and MRI measurements,²⁶ suggesting that our method’s precision is limited only by the modality itself rather than segmentation failure. Further, our error margin, including an absolute percent error of 5%, is also superior to inter- and intra-observer variability of MRI-based RV volumetry.^27,28 Therefore, RV measurements derived from our pipeline are clinically reliable.

DL algorithms have previously shown utility in 3D medical image segmentation. In CT, whole heart segmentation models^12,14 have been developed and RV segmentation accuracy has been assessed in patients undergoing TAVR.¹³ These models achieved RV Dice scores above 0.80 but used image patching to reduce memory demands.^12–14 Our 3D nnU-Net approach did use patching, which reduces the number of training examples, but simplifies processing. Alternatively, our group has shown that subsampling based on intensity regions improves processing time without sacrificing image and segmentation fidelity.¹¹ This requires preprocessing and network modification, which can be mitigated by using higher RAM GPUs.

The versatility of nnU-Net has enabled DL approaches to cardiac^16,29–31 and CT-based^30,32,33 imaging tasks. Efforts have focused on RV segmentation from short-axis,¹⁵ long-axis, and multi-view cardiac MRI.^16,29 Martín-Isla et al and Punithakumar et al similarly demonstrated that nnU-Net frameworks effectively predict RV segmentations across pathologies,^16,29 centres and scanners.¹⁶ They also noted that end-systolic segmentations were more error-prone than end-diastolic segmentations,¹⁶ a trend we also observed (Figure 3). Lastly, prior studies have largely reported model performance metrics like Dice, but have not directly assessed whether DL-derived segmentations enable accurate functional analysis.

PointNet-based architectures have been applied to cardiac imaging tasks, including 3D ventricular reconstruction from MRI,^34–36 prediction of cardiac deformations,³⁷ and detailed delineation of the LAA on CT³⁸. In our work, we use this architecture to automate labelling of the RV free and septal walls. Further work could subdivide these regions into AHA segments, though this would require a reliable method of manual labelling for training. Alternatively, rule-based subdivision (e.g. basal, mid, and apical) could be explored.

Differences in the extent of the RV labelled a particular wall did not significantly impact the accuracy of RV functional analysis; the accuracy of regional RV functional analysis was very strong, with a median FW and SW strain CSM of 0.97 in the fully automated segmentations.

We paired segmentation and labelling with volumetric and regional endocardial strain analysis. CT-based endocardial strain is reproducible³⁹ comparable to MR strain,^40,41 and robust to low-dose imaging.^20,41 CT-based RV strain has improved pre-operative risk stratification,⁹ matched clinical profiles of RV dysfunction,⁸ and has been combined with haemodynamics to estimate myocardial work.⁸ However, prior methods relied on time-consuming, expert CT image segmentations. While our group has established automated LV segmentation models,¹⁰ RV assessment has remained semi-automated. Here, we show that RHBS + RVWL can automate RV segmentation and yield RV strain profiles comparable to manual segmentation. This pipeline enables higher throughput and supports large-scale, multi-centre studies with improved reproducibility.

Given the diagnostic and prognostic utility of RV function in several cardiac disease pathophysiologies, our automated pipeline can be used to validate findings in small cohorts and establish new findings that were undetectable in underpowered cohorts. Namely, pre-procedural RV function has been shown to be linked to outcomes in patients undergoing TAVR,^42–46 as well as in LVAD recipients.^5,6,9 However, most studies measure RV functional globally,^{5,6,42–44,46} which overlooks local variations in function that we can capture with our pipeline. While some studies do evaluate regional RV function, these approaches are either conducted with echocardiography,⁴⁵ which can be unreliable in patients with dilated RVs, or in a relatively small cohort,⁹ which could warrant validation in a larger study. Additionally, procedural planning for TPVR in ACHD can also benefit from regional RV functional assessment by providing quantitative, regional information that can inform optimal time to intervention beyond current guidelines, which rely on symptoms and severe RV dilation.² RV dysfunction is also a key feature of CTEPH and can impact surgical outcomes.⁴⁷ However, quantitative assessment of the association between RV dysfunction severity and outcomes has yet to be evaluated.

Despite these strengths, our study has several limitations. First, it was conducted at a single centre using one CT scanner, and all manual segmentations for training were generated by a single expert. The lack of standardized training for RV segmentation from cineCT made broader manual labelling impractical. As a result, our automated pipeline learned the single observer’s segmentation style and may not be immediately generalizable to other readers. Future external validation should be conducted at multiple sites and with different CT vendors. Second, while the training cohort included patients with a range of complex anatomies, it lacked a truly ‘normal’ dataset, since patients with normal cardiac function rarely undergo CT. Third, although we validated our findings in an independent cohort with a separate aetiology from the training cohort, future work should test RHBS and RVWL in other populations. Fourth, while training cases spanned a clinical range of iodine enhancement, we selected TAVR case studies with varied enhancement to assess robustness (Figure 3). Fifth, while CSMs were largely >0.90, indicating that two strain histograms are very similar and therefore the assessment of function would be comparable, it remains to be studied to what extent accuracy is needed in specific categories, like the percent considered akinetic. This would depend on use cases. Lastly, the applicability of our combined approach (segmentation + wall labelling) to other chambers (e.g. left atrium) is unknown.

Conclusion

In this work, we present a fully automated pipeline to perform volumetry and regional strain analysis of the RV from contrast-enhanced, ECG-gated cineCT images. We developed and cross-validated two DL-based networks (RHBS and RVWL) in a diverse population of RV pathologies and then demonstrated clinical utility in an independent testing cohort.

Supplementary data

Supplementary data are available at European Heart Journal - Imaging Methods and Practice online.

Funding

This work was supported by National Institutes of Health (NIH) grants F31HL165881 (to A.C.), T32 HL 007444–43 (to A.C), and K01 HL143113 (to F.C.), and by a Schmidt AI in Science Postdoctoral Fellowship from the Eric and Wendy Schmidt Foundation (to L.S.).

Data availability

The data underlying this article will be shared on request to the corresponding author.

Lead author biography

Amanda Craine is a postdoctoral fellow in the Departments of Medicine and Bioengineering at the University of California, San Diego. Her research focuses on building tools to measure regional function from 4D CT and profiling right ventricular function in different pathophysiologies.