A computer-aided method for the automatic localization of the left ventricular long axis on coronary computed tomography angiography

Introduction

Given the persistently high mortality rates associated with cardiovascular diseases in various countries and regions (1), early screening and precise diagnosis are critical. Coronary computed tomography angiography (CCTA) is a noninvasive tool that allows for both visual and qualitative/quantitative analysis of the heart (2). The American Heart Association (AHA) recommends positioning and observing cardiac imaging modalities at a plane perpendicular to the long axis (defined as the line connecting the midpoint of the mitral valve to the apex) (3). However, computed tomography angiography (CTA) scanning is conventionally oriented along the vertical human body axis, neglecting the physiological relationship between the heart and the human body. Therefore, interpreting cardiac images derived from routine CT scan directions may result in incorrect localization of disease and affected areas. An anatomical position that aligns with the specific observational requirements is crucial for accurate disease diagnosis (4).

The long axis is the basis for determining the standard imaging plane of the heart, and the myocardial tissue structure can be evaluated in detail through multiple long-axis planes and one short-axis stack. As research into CCTA has intensified in recent years (5-7), post-processing of images derived from the long axis of the left ventricle has become a critical step in correcting for the observation angle deviation.

Some methods estimate the left ventricular long axis by fitting an ellipsoid model to the myocardial contour and using the major axis of the fitted ellipsoid as an approximation of the ventricular long axis. However, this method is limited in patients with morphological variations, as it cannot accurately fit the myocardial base, thereby compromising localization based on the true anatomical and physiological state (8,9). The use of a semiautomated approach to estimate positioning parameters on CCTA and to determine the orientation of the heart (10) involves human-computer interaction. The applicability of most methods to CCTA images remains limited, indicating the need to develop novel approaches specifically designed for CCTA imaging. With the rapid development of deep learning in cardiovascular imaging, myocardial segmentation on CCTA has become increasingly accurate and reproducible, enabling reliable anatomical localization and facilitating downstream quantitative analysis (11,12).

To develop a user-friendly approach, we constructed and validated an algorithm for left ventricular long-axis localization, which utilizes the unique anatomical features of the lower two-thirds of the myocardium. This algorithm computes and fits tangent lines to the outer edge of the myocardium, determining the cardiac long axis without requiring manual annotation. This approach streamlines the image post-processing workflow by eliminating interphysician variability and bias, increasing clinical utility and reproducibility. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2256/rc).

Methods

Study population

A total of 92 patients who underwent CTA examinations at two campuses of Beijing Friendship Hospital, Capital Medical University from December 2023 were consecutive enrolled. The exclusion criteria were (I) a history of valvular surgery and (II) CTA images showing misalignment or poor image quality. Consequently, four patients were excluded, three due to image artifacts and one due to mitral valve surgery. Thus, 88 patients were ultimately enrolled (46 males and 42 females). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (No. 2024-P2-009). All participants provided written informed consent to participate in the study.

CTA scanning protocol

CCTA data were acquired at two centers with either a 256-slice scanner (Revolution CT, GE HealthCare, Chicago, IL, USA) or a 320-slice scanner (Aquilion ONE, Canon Medical Systems, Otawara Japan) with prospective electrocardiography-gated acquisition. Heart rate control to approximately 60–75 beats/min was achieved through oral β-blocker administration when indicated. Iodinated contrast medium (iopamidol) was injected via a dual-barrel high-pressure injector at 4.0–5.5 mL/s for a total volume of 40–60 mL, which was followed by the administration of a 30-mL saline chaser at the same flow rate. Images were reconstructed with a slice thickness/spacing of 0.625/0.625 mm or 0.5/0.5 mm at 120 KV with automatic tube-current modulation and subsequently resampled to isotropic voxels. Reconstructions were generated at mid-diastole (~70–80% R-R interval), a routine clinical phase with minimal cardiac motion. No datasets were excluded based on heart rate or reconstructed phase to reflect real-world acquisition conditions.

Computer-based left ventricular long-axis localization

In the three-dimensional (3D) model of the myocardium, the left ventricular myocardium is approximately shaped like a cone with the apex of the heart, and when it is converted into a two-dimensional (2D) section (cross-section or sagittal plane), the myocardial contour appears as an isosceles triangle. Therefore, the angle bisector of the apex angle can be used to approximate the projection of the long axis on that specific plane.

Calculations were performed in the transverse and sagittal planes with good symmetry of the left ventricular myocardium through the following steps: (I) the computation included identifying the minimum bounding box that enclosed the entire myocardium. The rotation angle θ and rotation center C were used to perform a low-precision rotation on the 3D segmentation model of the left ventricle. This improved the axial length of the myocardium in the transverse and sagittal section; (II) for the 2D section, a specific area on the left and right borders was selected, and the parameters were adjusted so that the middle part of the myocardium that displayed significant symmetry for tangential calculation was used. The symmetry axis of the left and right tangential lines served as an approximate projection of the 3D long axis on the section plane, and it was then corrected; (III) the slice sequence was calculated one by one along the z-axis, the mean value of the long axes of each section was obtained, and the dimension along the z-axis was expanded to obtain the intermediate plane α1 and the intermediate plane α2, which were perpendicular to the xy plane and yz plane, respectively; (IV) the intersecting planes created an intersection line that represented the long axis of the left ventricle (Axis _d). Figure 1 illustrates the process for calculating the left ventricular long axis. The computation time was recorded as the total processing time per CCTA case on the tested workstation.

Figure 1 The process for obtaining the left ventricular long axis. C: rotation center; θ: rotation angle; brown dashed line: apex angle bisector; green dashed line: long-axis projection on the corresponding plane after correction; blue dashed line: left ventricular long axis. 2D, two-dimensional; 3D, three-dimensional.

Left ventricular long axis: physician annotation

All images were exported in Digital Imaging and Communications in Medicine (DICOM) format from the picture archiving and communication system (PACS) workstation. Two cardiovascular research-oriented physicians with over 3 years of experience used the open-source software 3D Slicer version 5.4.0 for manual editing on the original CTA images while being blinded to clinical information. The long axes labeled by different clinicians are denoted as Axis _a and Axis _b, respectively. The details of the workflow are provided in Appendix 1.

Data assessment: evaluation and processing

The processing of volume data in CTA involves an anatomical coordinate system (left-posterior-superior) and a geometric coordinate system (xyz). The conversion between these systems was completed through coordinate axis correspondence and proportional conversion. The details for the implementation of this conversion are described in Appendix 2.

With the unit direction vector Z^ along the z-axis serving as reference standard, relevant evaluation metrics were calculated for all the long axes in xyz. The angle α was defined as the rotation angle around the y axis required to align the intermediate vector α→ with the target axis vector tn→; β was defined as the rotation angle around the z-axis required to align with the target vector tn→ (Figure 2). The angle (γ) and distance (dist) between any two axes were calculated as shown in Eqs. [1] and [2], respectively.

γ=arccos(t1→⋅t2→)[1]

Dist=|(p1−p2)⋅t1→×t2→‖t1→×t2→‖|[2]

Figure 2 Illustration of α and β in the coordinate system. α: rotation angle around the y axis; β: rotation angle around the z axis; tn→: target axis vector; Z^: unit direction vector of the z-axis; α→: intermediate vector.

where (p1,t1→) and (p2,t2→) represent Axis _d and Axis _a, respectively; and the unit of α, β, and γ is the degree, while that of disy is a pixel.

Correction of the left ventricular long axis

During the visualization phase, Axis _d exhibited a more consistent deviation at the basal myocardium, leaning towards the direction of the interventricular septum when compared with Axis _a. Therefore, correction of the long axis direction was conducted to bring it closer to the center. The correction formula is shown in Eq. [3]:

t→=λ1t0→+λ2(xl+xr2−p)[3]

where x_l and x_r are the two vertices of the myocardial base, p is the intersection point obtained from Eq. [3], and λ1 and λ₂ are the correction coefficients satisfying λ₁ + λ₂ =1.

Statistical analysis

All analyses were performed in GraphPad Prism software version 9.0 (Dotmatics, Boston, MA, USA) or Python version 3.11 (Python Software Foundation, Wilmington, DE, USA). Normally distributed data are presented as the mean ± standard deviation or as the median and interquartile range (IQR). Categorical variables are reported as frequencies and percentages. Based on the data distribution, paired t-tests or Wilcoxon signed-rank tests were conducted for paired comparisons. Intraclass correlation coefficients (ICCs) and Bland-Altman were used to assess the agreement between the computer-predicted axis and physician annotations. Two radiologists independently labeled the long axis for evaluation of interreader agreement, while one radiologist repeated the labeling after a 3-week interval for evaluation of intrareader agreement. Spearman correlation was used to analyze the correlation between measurements obtained from the computer-predicted axis and physician annotation. A two-tailed P value <0.05 was considered statistically significant.

Results

Baseline characteristics

The baseline characteristics of the 88 study participants are summarized in Table 1. The mean age of the study population was 60.49±12.27 years, and 46 (52.3%) were male. A small proportion (n=23, 26.1%) had a history of coronary artery disease. Common atherosclerotic risk factors included diabetes mellitus (n=27, 30.7%), hypertension (n=56, 63.6%), and hyperlipidemia (n=62, 64.8%).

Correction

For the correction coefficient in Eq. [3], multiple sets of values were tested, with the most effective outcome being generated from λ1:λ₂=2:1. Typical cases are presented in Figure 3.

Figure 3 Visualization of the long-axis localization on 3D LV models reconstructed from CCTA. The yellow, red, blue, and green lines represent the Axis _d, Axis _a, and Axis _b, respectively. (A) A 68-year-old woman with hypertension and hyperlipidemia but not coronary artery disease. The α values were 100.42 and 101.89 for Axis _a and Axis _d, respectively; the β values were 134.62 and 133.04 for Axis _a and Axis _d, respectively. (B) A 62-year-old man with coronary artery disease, hyperlipidemia, and hypertension. The α values were 112.71 and 116.02 for Axis _a and Axis _d, respectively. The β values were 118.32 and 120.62 for Axis _a and Axis _d, respectively. 3D, three-dimensional; Axis _a, physician-annotated axis; Axis _b, second observer axis; Axis _d, computer-predicted axis; CCTA, coronary computed tomography angiography; LV, left ventricle; α, angle relative to the y axis; β, angle relative to the z axis.

Axis _a and Axis _d formed the following median angles with the coronal and transverse planes, respectively (Figure 4): α_a, 104.10° (IQR, 98.95–111.92°); α_d, 105.85° (IQR, 98.44–111.84°); β_a, 126.03° (IQR, 120.89–133.78°); and β_d, 125.57° (IQR, 121.38–133.66°). There was no significant significance in α or β between Axis _a and Axis _d (all P>0.05). The ICCs between radiologists were >0.9 for both α and β.

Figure 4 These box plots show the distribution of α and β obtained by different methods. α, angle relative to the y axis; β, angle relative to the z axis; NS, no significant difference.

Correlation and agreement

The algorithm required approximately 244 seconds (~4 minutes) to process a single CCTA case on a workstation equipped with an Intel Core i7-9700 CPU and 32 GB RAM on Windows 10 (64 bit). Spearman correlation analysis showed strong correlations between α measured from Axis _a and Axis _d (r=0.83, P<0.001) and between β measured from Axis _a and Axis _d (r=0.78, P<0.001). Scatterplots showed an approximately linear distribution of α and β at Axis _a and Axis _d (Figure 5).

Figure 5 Scatterplots showing the relationship between measurements derived from Axis _a and Axis _d for α and β. α, angle relative to the y axis; β, angle relative to the z axis. Axis _a, physician-annotated axis; Axis _d, computer-predicted axis.

Bland-Altman analysis revealed SDs of bias for α and β of 8.8 [95% limits of agreement (LoA): –16.98 to 17.65] and 4.62 (95% LoA: –9.49 to 8.60), respectively. The Bland-Altman plot (Figure 6) did not indicate any obvious systematic errors, suggesting that the differences between the Axis _a and Axis _d were randomly distributed and clinically acceptable.

Figure 6 The Bland-Altman plots showed no substantial systematic deviation between α and β measured at Axis _a and Axis _d. Each blue dot represents a piece of data. α, angle relative to the y axis; β, angle relative to the z axis; Axis _a, physician-annotated axis; Axis _d, computer-predicted axis; LoA, limits of agreement.

Relative differences

Generally, the angle error between the two axes was considered qualified within 10° (13). The median γ and the median dist between Axis _a and Axis _b were 3.84 (IQR, 2.13–6.22) and 2.85 (IQR, 1.22–4.48), respectively, whereas those between Axis _a and Axis _d were 4.40 (IQR, 2.36–8.66) and 19.42 (IQR, 10.25–36.7), respectively (Table 2). For the inter-observer comparison (Axis _a vs Axis _b), 90.9% of cases had an angular difference <10°, and 97.7% had a distance difference <20. For the method comparison (Axis _avs. Axis _d), the corresponding values were 82.9% and 52.3%, respectively.

Discussion

This study developed a long-axis localization algorithm for the left ventricle based on outer contour slopes derived from CCTA images which could capture individual anatomical characteristics. The results showed high agreement between the long axis determined by the algorithm and that annotated by clinicians. Further adjustments refined the visualization, allowing for more precise and clinically applicable long-axis localization, thereby providing a foundation for subsequent image post-processing and myocardial analysis.

The left ventricular long axis, the imaging plane used for cardiac imaging, has a certain impact on the accuracy of the calculation of parameters such as left ventricular volume and ejection fraction (13,14). Some methods use deep learning to identify certain key points within the heart, enabling the determination of commonly used scanning planes (15,16). Other approaches directly employ models to compute imaging planes on CCTA (17). However, there remains a need for physicians to annotate images as a training set, and this annotation process can be time-consuming, while the training and computation may be complex. In our study, we directly extracted left ventricular myocardial data from routine CCTA images and utilized the true anatomical structure as prior knowledge, avoiding the limitations of single geometric models, such as conical- or bullet-shaped assumptions, which disregard individual cardiac morphology. During the study, no manual preprocessing or annotation of the training set by clinicians was required, reducing both inter- and intraobserver variability. In this approach, the long-axis generation is fully automated, offering rapid and highly reproducible results that are suitable for routine clinical diagnostic use. The differences between the Axis _a and Axis _d values predominantly fell within a range of less than 10°, meeting the clinical diagnostic requirements.

Lelieveldt et al. (18) reported a median relative angle γ of 12.2°±6.8° based on cardiac magnetic resonance, whereas our CCTA-based analysis yielded a lower value of 4.40°. Although the two modalities differ in contrast use and spatial resolution, γ primarily reflects myocardial segment geometry rather than signal intensity, and both studies reconstructed images in mid-diastole, minimizing motion and ensuring comparable geometric representation. The processing time in our study was approximately 244 seconds (~4 min) per case, while that in Lelieveldt et al.’s study was 5–7 minutes. Differences in hardware, software, and technological era limit direct comparison, yet these observations provide a useful benchmark for the efficiency of modern CCTA-based methods. Overall, geometry-derived descriptors of LV anatomy appear inherently stable, showing reproducibility across imaging modalities and computational frameworks, as well as insensitivity to demographic and physiological factors, including BMI, age, respiratory motion, and cardiac cycle. This combination of technical and physiological robustness underscores the suitability of geometry-based metrics for reliable long-axis localization and comparative myocardial analysis (19,20).

The agreement for α was slightly inferior to that of β, although both demonstrated strong correlation. This may be primarily attributed to the complexity of the anatomical structure at the basal myocardium. To address this issue, a correction method was applied to reduce the deviation of the long axis at the basal myocardium. This method adjusts the long axis toward the center at the base, reducing the relative distance differences between axes. The results indicated that the impact of this offset error can be disregarded.

This study involved several limitations that should be acknowledged. First, no patients with significant ventricular abnormalities (e.g., distorted or D-shaped ventricles) were included, and future work could include patients to improve the robustness and generalizability of the model. Second, due to the retrospective design and limited sample size, expansion of the cohort is needed to enhance reliability and external validity. Finally, the influence of clinical characteristics on long-axis localization was not analyzed, and potential variability across cardiac conditions remained to be further investigated.

Conclusions

This study established a left ventricular long-axis localization method based on myocardial segmentation, in which the outer contour slopes are calculated to align with anatomical features. The method demonstrated accuracy and consistency comparable to those of clinicians while significantly reducing image post-processing time. This approach can be integrated into routine cardiac imaging workflows to enable high-precision lesion localization.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2256/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2256/dss

Funding: This research was funded by the National Natural Science Foundation of China (No. 82272068) and Beijing Municipal Natural Science Foundation (No. L256013).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-aw-2256/coif). Shuang Li and Y.H. report funding support from National Natural Science Foundation of China (No. 82272068) and Beijing Municipal Natural Science Foundation (No. L256013). Yinghao Xu is an employee of Canon Medical Systems (China) Co., Ltd. and the company has no conflicts of interest related to this study. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Friendship Hospital, Capital Medical University (No. 2024-P2-009). All participants provided written informed consent to participate in the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Li Y, Cao GY, Jing WZ, Liu J, Liu M. Global trends and regional differences in incidence and mortality of cardiovascular disease, 1990-2019: findings from 2019 global burden of disease study. Eur J Prev Cardiol 2023;30:276-86. [Crossref] [PubMed]
2. Abdelrahman KM, Chen MY, Dey AK, Virmani R, Finn AV, Khamis RY, et al. Coronary Computed Tomography Angiography From Clinical Uses to Emerging Technologies: JACC State-of-the-Art Review. J Am Coll Cardiol 2020;76:1226-43. [Crossref] [PubMed]
3. Cerqueira MD, Weissman NJ, Dilsizian V, Jacobs AK, Kaul S, Laskey WK, Pennell DJ, Rumberger JA, Ryan T, Verani MSAmerican Heart Association Writing Group on Myocardial Segmentation and Registration for Cardiac Imaging. Standardized myocardial segmentation and nomenclature for tomographic imaging of the heart. A statement for healthcare professionals from the Cardiac Imaging Committee of the Council on Clinical Cardiology of the American Heart Association. Circulation 2002;105:539-42. [Crossref] [PubMed]
4. Squara F, Fourrier E, Diascorn Y, Poulard A, Scarlatti D, Bun SS, Moceri P, Ferrari E. Cardiac anatomical axes by CT scan and confirmation of the accuracy of fluoroscopic individualized left anterior oblique projection for right ventricular lead implantation. J Interv Card Electrophysiol 2021;60:213-9. [Crossref] [PubMed]
5. Thomsen C, Abdulla J. Characteristics of high-risk coronary plaques identified by computed tomographic angiography and associated prognosis: a systematic review and meta-analysis. Eur Heart J Cardiovasc Imaging 2016;17:120-9. [Crossref] [PubMed]
6. Polidori T, De Santis D, Rucci C, Tremamunno G, Piccinni G, Pugliese L, Zerunian M, Guido G, Pucciarelli F, Bracci B, Polici M, Laghi A, Caruso D. Radiomics applications in cardiac imaging: a comprehensive review. Radiol Med 2023;128:922-33. [Crossref] [PubMed]
7. Aromiwura AA, Settle T, Umer M, Joshi J, Shotwell M, Mattumpuram J, Vorla M, Sztukowska M, Contractor S, Amini A, Kalra DK. Artificial intelligence in cardiac computed tomography. Prog Cardiovasc Dis 2023;81:54-77. [Crossref] [PubMed]
8. deKemp RA, Nahmias C. Automated determination of the left ventricular long axis in cardiac positron tomography. Physiol Meas 1996;17:95-108. [Crossref] [PubMed]
9. Jackson CE, Robson MD, Francis JM, Noble JA. Computerised planning of the acquisition of cardiac MR images. Comput Med Imaging Graph 2004;28:411-8. [Crossref] [PubMed]
10. Park S, Chung M. Cardiac segmentation on CT Images through shape-aware contour attentions. Comput Biol Med 2022;147:105782. [Crossref] [PubMed]
11. Jun Guo B, He X, Lei Y, Harms J, Wang T, Curran WJ, Liu T, Jiang Zhang L, Yang X. Automated left ventricular myocardium segmentation using 3D deeply supervised attention U-net for coronary computed tomography angiography; CT myocardium segmentation. Med Phys 2020;47:1775-85. [Crossref] [PubMed]
12. Guo Y, Bi L, Zhu Z, Feng DD, Zhang R, Wang Q, Kim J. Automatic left ventricular cavity segmentation via deep spatial sequential network in 4D computed tomography. Comput Med Imaging Graph 2021;91:101952. [Crossref] [PubMed]
13. Nosir YF, Vletter WB, Boersma E, Frowijn R, Ten Cate FJ, Fioretti PM, Roelandt JR. The apical long-axis rather than the two-chamber view should be used in combination with the four-chamber view for accurate assessment of left ventricular volumes and function. Eur Heart J 1997;18:1175-85. [Crossref] [PubMed]
14. Mullick R, Ezquerra NF. Automatic determination of LV orientation from SPECT data. IEEE Trans Med Imaging 1995;14:88-99. [Crossref] [PubMed]
15. Blansit K, Retson T, Masutani E, Bahrami N, Hsiao A. Deep Learning-based Prescription of Cardiac MRI Planes. Radiol Artif Intell 2019;1:e180069. [Crossref] [PubMed]
16. Xue H, Artico J, Fontana M, Moon JC, Davies RH, Kellman P. Landmark Detection in Cardiac MRI by Using a Convolutional Neural Network. Radiol Artif Intell 2021;3:e200197. [Crossref] [PubMed]
17. Chen Z, Rigolli M, Vigneault DM, Kligerman S, Hahn L, Narezkina A, Craine A, Lowe K, Contijoch F. Automated cardiac volume assessment and cardiac long- and short-axis imaging plane prediction from electrocardiogram-gated computed tomography volumes enabled by deep learning. Eur Heart J Digit Health 2021;2:311-22. [Crossref] [PubMed]
18. Lelieveldt BP, van der Geest RJ, Lamb HJ, Kayser HW, Reiber JH. Automated observer-independent acquisition of cardiac short-axis MR images: a pilot study. Radiology 2001;221:537-42. [Crossref] [PubMed]
19. Engblom H, Hedström E, Palmer J, Wagner GS, Arheden H. Determination of the left ventricular long-axis orientation from a single short-axis MR image: relation to BMI and age. Clin Physiol Funct Imaging 2004;24:310-5. [Crossref] [PubMed]
20. Foster JE, Engblom H, Martin TN, Wagner GS, Steedman T, Ferrua S, Elliott AT, Dargie HJ, Groenning BA. Determination of left ventricular long-axis orientation using MRI: changes during the respiratory and cardiac cycles in normal and diseased subjects. Clin Physiol Funct Imaging 2005;25:286-92. [Crossref] [PubMed]