Abstract
Background
Cardiac magnetic resonance (CMR) is the reference standard for evaluating left ventricular (LV) function and structure. Manual post-processing is time-consuming, operator-dependent, and limited in reproducibility. Deep learning-based artificial intelligence (AI-DL) offers automated image segmentation with expert-level accuracy, but requires clinical validation against manual analysis. This study aimed to directly compare two AI-DL segmentation strategies—simplified (trabeculae and papillary muscles included in blood pool) and standard (included in myocardium)—with manual analysis to determine which approach yields superior clinical agreement.
Methods
We retrospectively studied 104 patients who underwent CMR. Using both AI-DL (simplified/standard) methods and traditional manual post-processing software, we quantified LV parameters including ejection fraction (EF), end-diastolic volume and index (EDV/EDVI), end-systolic volume and index (ESV/ESVI), stroke volume and index (SV/SI), cardiac output and index (CO/CI), and end-diastolic LV-mass and index (M/MI). Analysis time, correlation, and agreement with manual results were compared.
Results
AI-DL markedly reduced analysis time compared with manual post-processing following standard method (median 50.0 vs. 217.0 s; P<0.001). The AI-DL simplified method underestimated EF (54.2 vs. 60.0), mass indices (131.45 vs. 156.00, 76.45 vs. 91.25) and overestimated volumetric indices (148.10 vs. 118.00, 88.70 vs. 69.80, 62.85 vs. 46.65, 40.00 vs. 28.70, 69.90 vs. 63.30, 41.70 vs. 37.83, 5.35 vs. 4.80, 3.15 vs. 2.80) relative to manual analysis (all P<0.001). In contrast, the standard method showed no significant differences from manual values (58.3 vs. 60.0, 122.80 vs. 118.00, 72.70 vs. 69.80, 46.30 vs. 46.65, 27.75 vs. 28.70, 64.70 vs. 63.30, 37.85 vs. 37.83, 4.90 vs. 4.80, 2.90 vs. 2.80, 163.05 vs. 156.00, 93.55 vs. 91.25) (all P>0.05). Both methods demonstrated strong correlations with manual results, with the standard method achieving higher intraclass correlation coefficient (0.798–0.978 vs. 0.690–0.928) and correlation coefficients (r=0.772–0.952 vs. 0.726–0.944) (all P<0.001) as well as narrower Bland-Altman limits of agreement (LOA) (the differences were smaller and not statistically significant, all P>0.05), whereas the simplified method had wider LOA (the differences were larger and statistically significant, all P<0.001).
Conclusions
AI-DL enables rapid and reliable quantification of LV function and structure in CMR. Although both simplified and standard methods show strong agreement with manual analysis, the standard method provides better agreement and narrower limits of difference, making it more clinically acceptable. These findings highlight the potential of AI-DL as a practical tool for routine cardiac functional assessment.
Keywords: Artificial intelligence (AI), deep learning, cardiac magnetic resonance (CMR), function and structure parameters, comparison
Introduction
Cardiovascular disease (CVD) has remained the leading cause of morbidity and mortality in the world over the past few decades (1,2). Targeted and sensitive diagnostic methods are essential for early diagnosis and prevention. For left ventricular (LV) assessment, accurate evaluation of structural and functional parameters, including myocardial mass, volume, and ejection fraction (EF) is critical for accurate diagnosis, prognosis, and therapeutic decision-making in CVD (3). Therefore, a rapid and reliable cardiac examination of functional and structural parameters is of great significance for disease diagnosis, treatment, follow-up evaluation, and risk estimation (4).
Cardiac magnetic resonance (CMR) provides high-resolution and multi-parametric assessment of cardiac function and morphology, and is regarded as the gold standard for LV evaluation (5). Compared with CMR, other imaging methods such as computed tomography (CT) and ultrasound have inherent limitations in image quality and safety (6,7). Cine CMR enables accurate quantification of LV functional and structural parameters without the need for gadolinium contrast. However, conventional post-processing relies on manual contour delineation, which is labor-intensive and prone to inter-observer variability, particularly in basal and apical slices where the anatomy is often ambiguous (8). These limitations have motivated the development of automated segmentation methods based on artificial intelligence (AI), particularly deep learning (DL), which can accelerate analysis and reduce operator dependency (9,10).
Recent advances in cine short-axis (SAX) segmentation algorithms have introduced classifier-guided two-stage networks combined with all-slice fusion transformers. This framework mimics the radiologist’s annotation workflow by first classifying anatomical presence and then performing standard segmentation, thereby integrating both intra- and inter-slice contextual information. These methods address the shortcomings of traditional single-stage approaches, which often struggle with complex cardiac morphology and yield suboptimal performance in challenging slices (11). By improving contour accuracy and visual consistency—reducing artifacts such as holes or fragmented regions—cine SAX algorithms enable more robust and reliable quantification of LV mass, volumes, and EF without the need for contrast agents. This progress not only enhances the accuracy of AI-DL post-processing but also facilitates broader clinical translation of automated CMR analysis.
Despite these developments, evidence remains limited regarding direct comparisons between AI-DL methods that differ in their treatment of trabeculae and papillary muscles (TPM). The simplified method assigns TPM to the blood pool, whereas the standard method assigns TPM to the myocardium (12,13). Determining the optimal AI-DL strategy therefore requires validation against manual post-processing to assess which approach—simplified or standard—achieves better agreement with expert analysis. However, few studies to date have compared the performance of AI-DL for cardiac functional structural analysis between the simplified and standard CMR methods.
In this study, we retrospectively analyzed clinical CMR data from patients and compared LV functional and structural parameters obtained from AI-DL (simplified and standard methods) with those from manual post-processing, with the aim of evaluating the clinical applicability of AI-DL for CMR analysis. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0364/rc).
Methods
Study design and population
A single-center, retrospective study was conducted in the Department of Radiology, The Third Affiliated Hospital of Sun Yat-sen University Yuedong Hospital. A total of 120 patients who underwent 3.0 Tesla (T) CMR scans from February 2024 to August 2025 were screened. The inclusion criteria were as follows: (I) clear consciousness with the ability to cooperate during CMR scanning; and (II) cine images of sufficient image quality to delineate endocardial and epicardial contours. The exclusion criteria were as follows: (I) inability to cooperate due to claustrophobia; and (II) patients with pacemakers and aneurysm clips in their bodies. A total of 104 patients were ultimately included in the study. Figure 1 shows a detailed flowchart illustrating the above process. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The Third Affiliated Hospital of Sun Yat-sen University Yuedong Hospital (No. II2026-004-01) and all patients provided written informed consent.
Figure 1.
Patient selection flowchart. CMR, cardiac magnetic resonance.
Magnetic resonance imaging acquisition
Imaging was performed either on a Philips Ingenia Elition 3.0 T scanner (Philips, Heerlen, Netherlands) with an 18-channel phased-array body coil or a GE SIGNA Architect 3.0T (GE Healthcare, Milwaukee, WI, USA) with a flexible 30-channel phased-array body coil. Among the 104 included patients, 90 (86.5%) were examined on the Philips Ingenia Elition 3.0T and 14 (13.5%) on the GE SIGNA Architect 3.0T. Patients were scanned in the supine position using electrocardiogram (ECG) gating with breath-hold instruction. Sequences were acquired sequentially in the following order: a three-plane localizer (axial, coronal, sagittal), a black-blood sequence covering the region from the aortic arch to the diaphragmatic surface, and one or more sets of 2D SSH BIFE sequences at the same anatomical location. Subsequently, two-chamber (2ch), four-chamber (4ch), and a stack of short-axis images (10–15 slices) of the left ventricle were obtained. The typical imaging parameters were as follows: repetition time (TR) 3.0–3.1 ms; echo time (TE) 1.0–1.5 ms; flip angle (FA) 40–50°; field of view (FOV) 300–400 mm2; matrix (172–180) × (180–288); slice thickness 8 mm; slice gap spacing 1–2 mm; pixel size ~1.8×1.6 mm2.
Manual post-processing
The cine images were analyzed on vendor-provided Philips IntelliSpace IX/LX Workstations (No. 105554601) or GE Advanced Workstation (version 4.7) by two radiologists (each with >5 years of CMR experience) who independently performed semi-automated segmentation using vendor software, followed by manual adjustment. Any discrepancies were resolved by consensus. Readers were blinded to all AI-DL outputs during manual analysis. Semi-automatic segmentation followed by manual correction was performed for the LV endocardium and epicardium at end-systole and end-diastole on short-axis stacks with TPM assigned to the myocardium by semiautomatic threshold selection using the vendors’ cardiac function analysis software (Figures 2,3). The following LV parameters were obtained: EF, end-diastolic volume and index (EDV/EDVI), end-systolic volume and index (ESV/ESVI), stroke volume and index (SV/SI), cardiac output and index (CO/CI), and mass and index (M/MI). The analysis time for each case was recorded.
Figure 2.
Left ventricular structural and functional analysis in a 17-year-old male presenting with syncope due to chest tightness. (A,B) The manual method: manual delineation of the short-axis slice showing the endocardial (green line) and epicardial (yellow line) contours with derived left ventricular parameters. The TPM (blue) was attributed to the myocardium. (C,D) The AI-DL simplified method: automatic segmentation of endocardial and epicardial contours on the short-axis slice with derived left ventricular parameters and the TPM was assigned to the blood pool. (E,F) The AI-DL standard method: automatic segmentation of endocardial and epicardial contours with derived left ventricular parameters and the TPM (yellow line) was attributed to the myocardium. AI, artificial intelligence; BSA, body surface area; DL, deep learning; ED, end-diastolic; EDV, end-diastolic volume; ES, end systolic; LV, left ventricular; N/A, not applicable; TPM, trabeculae and papillary muscles.
Figure 3.
Left ventricular structural and functional analysis in a 54-year-old female with coronary heart disease. (A,B) The manual method: manual delineation of the short-axis slice showing the endocardial (red line) and epicardial (green line) contours with derived left ventricular parameters. The TPM was attributed to the myocardium. (C,D) The AI-DL simplified method: automatic segmentation of endocardial and epicardial contours on the short-axis slice with derived left ventricular parameters and the TPM was assigned to the blood pool. (E,F) The AI-DL standard method: automatic segmentation of endocardial and epicardial contours with derived left ventricular parameters and the TPM (yellow line) was attributed to the myocardium. bpm, beats per minute; ED, end-diastolic; EDV, end-diastolic volume; ES, end systolic; LV, left ventricular; TPM, trabeculae and papillary muscles.
AI-DL analysis (simplified and standard methods)
The same cine image was processed using United Imaging Intelligent CMR Image Processing Software (United Imaging Intelligent Technology Co., Ltd., Shanghai, China. Version 20250130sp1), which performs automatic LV segmentation and quantitative analysis based on DL algorithms. The AI-DL segmentation pipeline was fully automated without any user intervention. End-diastolic and end-systolic phases were automatically identified as the first phase after the ECG R-wave trigger and the phase with minimum LV cavity area on the volume-time curve, respectively. Phase selection was visually confirmed to match the phases used in the clinical reference analysis for each case. The analysis time was recorded.
The AI-DL pipeline produced two sets of results depending on the treatment of TPM: the simplified method (Figures 2C,2D,3C,3D) with TPM assigned to the blood pool; the standard method (Figures 2E,2F,3E,3F) with TPM assigned to the myocardium, consistent with contemporary post-processing consensus (12,13). Phase selection was visually confirmed to match the phases used in the clinical reference analysis for each case. Any discrepancy in phase selection (observed in 17.3% of cases) was manually corrected prior to comparative analysis to avoid bias.
Endpoints and measurements
Primary endpoints were between-method differences, agreement, and correlations for left ventricular ejection fraction (LVEF), EDV/EDVI, ESV/ESVI, SV/SI, CO/CI, and M/MI among manual, AI-DL simplified, and AI-DL standard outputs. Analysis time served as a secondary endpoint to quantify workflow efficiency.
Statistical analysis
Statistical analyses were performed using SPSS 25.0 (IBM Corp., Armonk, NY, USA) and MedCalc 20.022 (MedCalc Software, Ostend, Belgium). The Shapiro-Wilk test was applied to assess the normality of continuous variables. The AI analysis time was compared with the manual analysis time. The parameters of LV function and structure calculated based on AI-DL (simplified method/standard method) and manual method were compared to analyze the difference, consistency, and correlation between the two methods. Normally distributed variables were expressed as mean ± standard deviation (SD) and compared using paired-sample t-tests. For skewed continuous variables, data were expressed as median [interquartile range (IQR)], and paired comparisons were performed using the Wilcoxon signed-rank test. Correlation analyses were conducted using Pearson correlation for normally distributed data and Spearman correlation for non-normally distributed data. Agreement between methods was evaluated by intraclass correlation coefficient (ICC) and Bland-Altman analysis. Correlation strength was defined as follows: r≤0.2, weak; 0.2<r≤0.4, mild; 0.4<r≤0.6, moderate; 0.6<r≤0.8, strong; and 0.8<r≤1.0, excellent. ICC was interpreted as ≤0.40, poor; 0.40<r≤0.75, moderate; and >0.75, good. For all statistical analysis, P<0.05 was considered statistically significant.
Results
Patient characteristics
A total of 104 patients who underwent CMR were included, comprising 80 males (76.9%) and 24 females (23.1%), with a median age of 57.5 years (IQR, 47.5–66.0 years). The mean body mass index (BMI) was 24.5±3.82 kg/m2, the median heart rate was 75 beats/min (IQR, 66–87 beats/min), and the mean body surface area (BSA) was 1.74±0.19 m2. The CMR diagnosis included: atrial septal defect (n=2, 1.9%), hypertrophic cardiomyopathy (n=3, 2.9%), hypertensive heart disease (n=6, 5.8%), coronary heart disease (n=57, 54.8%), dilated cardiomyopathy (n=14, 13.5%), myocarditis (n=1, 0.9%), arrhythmia (n=4, 3.8%), valvular disease (n=3, 2.9%), heart failure (n=3, 2.9%), and others with no significant abnormality (n=11, 10.6%).
Inter-observer reproducibility of manual analysis
To assess inter-observer variability, we randomly selected 20 patients (10 with normal LV dimensions, 10 with dilated ventricles) and compared the initial independent measurements of the two radiologists before consensus. The ICC were: EF =0.930 [95% confidence interval (CI): 0.834–0.971], EDV =0.931 (95% CI: 0.835–0.972), ESV =0.948 (95% CI: 0.876–0.979), CO =0.773 (95% CI: 0.510–0.904), and SV =0.853 (95% CI: 0.668–0.939), all indicating good reproducibility.
Analysis time
The median analysis time was significantly shorter for AI-DL than that of the manual method as listed in Table 1. The AI-DL methods analyzed in 50.0 seconds (IQR, 49.0–52.0 seconds), whereas the manual method needed 217.0 seconds (IQR, 211.5–223.0 seconds). This difference was statistically significant (Z=–8.855, P<0.001), demonstrating markedly improved efficiency with AI-DL.
Table 1. Analysis time comparison of AI-DL and manual methods.
| Methods | Value | Z | P value |
|---|---|---|---|
| AI-DL (seconds) | 50.0 (49.0, 52.0) | –8.855 | <0.001* |
| Manual (seconds) | 217.0 (211.5, 223.0) |
Data are presented as median (interquartile range). *, P<0.05 was considered statistically significant. AI, artificial intelligence; DL, deep learning.
LV functional and structural parameters
LV parameters calculated by manual, AI-DL simplified and AI-DL standard methods are summarized in Table 2.
Table 2. Left ventricular parameters calculated by manual and AI-DL methods.
| Left ventricular parameters | Manual | AI-DL simplified | AI-DL standard |
|---|---|---|---|
| EF (%) | 60.0 (34.5, 68.0) | 54.2 (28.4, 63.8) | 58.3 (33.4, 71.7) |
| EDV (mL) | 118.00 (97.70, 170.20) | 148.10 (122.80, 224.65) | 122.80 (96.00, 178.90) |
| EDVI (mL/m2) | 69.80 (57.25, 94.75) | 88.70 (70.95, 123.45) | 72.70 (57.95, 99.40) |
| ESV (mL) | 46.65 (32.70,100.00) | 62.85 (47.50,138.05) | 46.30 (29.65, 109.45) |
| ESVI (mL/m2) | 28.70 (18.70, 61.90) | 40.00 (27.15, 80.80) | 27.75 (17.95, 60.80) |
| SV (mL) | 63.30 (53.85, 76.55) | 69.90 (54.30, 87.85) | 64.70 (46.00, 78.00) |
| SI (mL/m2) | 37.83±12.42 | 41.70 (31.05, 52.80) | 37.85 (26.70, 47.45) |
| CO (L/min) | 4.80 (4.00, 5.60) | 5.35 (4.25, 6.55) | 4.90 (3.80, 5.80) |
| CI (L/min/m2) | 2.80 (2.30, 3.30) | 3.15 (2.50, 3.70) | 2.90 (2.20, 3.30) |
| M (g) | 156.00 (133.40, 195.88) | 131.45 (110.35, 158.70) | 163.05 (134.90, 202.28) |
| MI (g/m2) | 91.25 (76.55, 117.68) | 76.45 (62.53, 91.35) | 93.55 (78.58, 119.23) |
Data are presented as median (interquartile range) or mean ± standard deviation. AI, artificial intelligence; CI, cardiac index; CO, cardiac output; DL, deep learning; EDV, end-diastolic volume; EDVI, end-diastolic volume index; EF, ejection fraction; ESV, end-systolic volume; ESVI, end-systolic volume index; M, mass; MI, mass index; SI, stroke index; SV, stroke volume.
Phase selection for end-diastole and end-systole was identical between the AI-DL and manual methods in 86/104 cases (82.7%). In the remaining 18 cases (17.3%), end-systolic phase mismatch (±1 phase) was identified and manually corrected to ensure data consistency before quantitative comparison.
Comparison, consistency, correlation, and agreement analysis
The AI-DL simplified method tended to underestimate EF, M/MI, and overestimate EDV/EDVI, ESV/ESVI, SV/SI, and CO/CI compared with manual measurements that attributed TPM to myocardium (all P<0.001). The AI-DL standard method showed no significant differences from manual measurements in EF, EDV/EDVI, ESV/ESVI, SV/SI, CO/CI, or M/MI (all P>0.05), suggesting closer agreement in Table 3. Since the manual reference employed the standard method (TPM assigned to myocardium), the AI-DL simplified method was compared to manual results primarily to illustrate the expected directional bias arising from different TPM handling strategies, rather than to assess its absolute accuracy.
Table 3. The results of difference (Z), consistency (ICC), and correlation (Spearman) analyses of LV functional and structural parameters calculated by AI-DL (simplified/standard) and manual methods.
| LV functional and structural parameters | Between-group comparison | Z | P value | ICC (95% CI) | P value | r | P value |
|---|---|---|---|---|---|---|---|
| EF (%) | AI1 vs. Manual | −6.740 | <0.001 | 0.928 (0.754–0.968) | <0.001 | 0.926 | <0.001 |
| AI2 vs. Manual | −0.536 | 0.592 | 0.951 (0.928–0.966) | <0.001 | 0.931 | <0.001 | |
| EDV (mL) | AI1 vs. Manual | −8.610 | <0.001 | 0.857 (0.093–0.955) | <0.001 | 0.926 | <0.001 |
| AI2 vs. Manual | −0.547 | 0.584 | 0.963 (0.946–0.975) | <0.001 | 0.927 | <0.001 | |
| EDVI (mL/m2) | AI1 vs. Manual | −8.654 | <0.001 | 0.839 (0.072–0.949) | <0.001 | 0.921 | <0.001 |
| AI2 vs. Manual | −0.730 | 0.465 | 0.954 (0.933–0.969) | <0.001 | 0.925 | <0.001 | |
| ESV (mL) | AI1 vs. Manual | −7.674 | <0.001 | 0.901 (0.577–0.961) | <0.001 | 0.926 | <0.001 |
| AI2 vs. Manual | −0.188 | 0.851 | 0.958 (0.939–0.971) | <0.001 | 0.936 | <0.001 | |
| ESVI (mL/m2) | AI1 vs. Manual | −8.166 | <0.001 | 0.902 (0.407–0.966) | <0.001 | 0.944 | <0.001 |
| AI2 vs. Manual | −0.337 | 0.736 | 0.978 (0.968–0.985) | <0.001 | 0.952 | <0.001 | |
| SV (mL) | AI1 vs. Manual | −5.654 | <0.001 | 0.770 (0.568–0.867) | <0.001 | 0.826 | <0.001 |
| AI2 vs. Manual | −0.610 | 0.542 | 0.825 (0.752–0.878) | <0.001 | 0.844 | <0.001 | |
| SI (mL/m2) | AI1 vs. Manual | −5.588 | <0.001 | 0.708 (0.494–0.824) | <0.001 | 0.804 | <0.001 |
| AI2 vs. Manual | −1.171 | 0.242 | 0.798 (0.716–0.858) | <0.001 | 0.845 | <0.001 | |
| CO (L/min) | AI1 vs. Manual | −5.244 | <0.001 | 0.704 (0.507–0.816) | <0.001 | 0.759 | <0.001 |
| AI2 vs. Manual | −0.110 | 0.912 | 0.816 (0.740–0.871) | <0.001 | 0.796 | <0.001 | |
| CI (L/min/m2) | AI1 vs. Manual | −5.320 | <0.001 | 0.690 (0.475–0.810) | <0.001 | 0.726 | <0.001 |
| AI2 vs. Manual | −0.393 | 0.695 | 0.806 (0.727–0.864) | <0.001 | 0.772 | <0.001 | |
| M (g) | AI1 vs. Manual | −8.522 | <0.001 | 0.874 (0.820–0.913) | <0.001 | 0.917 | <0.001 |
| AI2 vs. Manual | −0.480 | 0.631 | 0.918 (0.881–0.944) | <0.001 | 0.931 | <0.001 | |
| MI (g/m2) | AI1 vs. Manual | −8.532 | <0.001 | 0.839 (0.771–0.888) | <0.001 | 0.904 | <0.001 |
| AI2 vs. Manual | −0.611 | 0.541 | 0.903 (0.861–0.934) | <0.001 | 0.910 | <0.001 |
AI1 is the AI-DL simplified method; AI2 is the AI-DL standard method; Manual is the manual method. AI, artificial intelligence; CI, cardiac index; CO, cardiac output; DL, deep learning; EDV, end-diastolic volume; EDVI, end-diastolic volume index; EF, ejection fraction; ESV, end-systolic volume; ESVI, end-systolic volume index; ICC, intraclass correlation coefficient; LV, left ventricular; M, mass; MI, mass index; SI, stroke index; SV, stroke volume.
Correlation and agreement analyses are summarized in Table 3. Both AI-DL methods demonstrated strong to excellent correlations (r=0.726–0.952) and moderate-to-good agreement (ICC =0.690–0.978) with manual analysis. The standard method consistently achieved higher ICC and correlation values than the simplified method, with narrower 95% limits of agreement (LOA).
Bland-Altman analysis further confirmed these findings (Figure 4). The simplified method systematically underestimated EF and overestimated volumetric indices, whereas the standard method demonstrated minimal bias and tighter LOA, indicating superior reliability in Table 4.
Figure 4.
Bland-Altman plots of AI-DL (simplified/standard) versus manual calculation of LV functional structural parameters. (A-V) The left ventricular EF (A,B), EDV (C,D), EDVI (E,F), ESV (G,H), ESVI (I,J), SV (K,L), SI (M,N), CO (O,P), CI (Q,R), M (S,T), and MI (U,V) calculated by AI1 and AI2, respectively, and the corresponding Bland-Altman plots of the left ventricular functional structural parameters calculated by manual method (note: AI1 is the AI-DL simplified method, AI2 is the AI-DL standard method, Manual is the manual method). AI, artificial intelligence; CI, cardiac index; CO, cardiac output; DL, deep learning; EDV, end-diastolic volume; EDVI, end-diastolic volume index; EF, ejection fraction; ESV, end-systolic volume; ESVI, end-systolic volume index; LV, left ventricular; M, mass; MI, mass index; SI, stroke index; SV, stroke volume.
Table 4. Differences and 95% LOA of Bland-Altman analysis of LV functional and structural parameters calculated by AI-DL (simplified/standard) and manual method.
| LV CMR functional and structural parameters | Between-group comparison |
Difference | P value | 95% LOA | |
|---|---|---|---|---|---|
| Mean | 95% CI | ||||
| EF (%) | AI1 vs. Manual | −4.723 | −5.939 to −3.507 | <0.001 | −16.974 to 7.529 |
| AI2 vs. Manual | 0.376 | −0.923 to 1.675 | 0.567 | −12.713 to 13.465 | |
| EDV (mL) | AI1 vs. Manual | 33.237 | 27.403 to 39.071 | <0.001 | −14.482 to 82.955 |
| AI2 vs. Manual | −0.630 | −3.746 to 3.619 | 0.973 | −37.181 to 37.054 | |
| EDVI (mL/m2) | AI1 vs. Manual | 19.984 | 17.141 to 22.827 | <0.001 | −8.668 to 48.636 |
| AI2 vs. Manual | −0.463 | −1.796 to 2.721 | 0.685 | −22.297 to 23.222 | |
| ESV (mL) | AI1 vs. Manual | 23.450 | 17.604 to 29.295 | <0.001 | −31.561 to 78.460 |
| AI2 vs. Manual | −2.710 | −6.715 to 1.296 | 0.183 | −43.082 to 37.662 | |
| ESVI (mL/m2) | AI1 vs. Manual | 14.976 | 12.302 to 17.650 | <0.001 | −11.970 to 41.921 |
| AI2 vs. Manual | −0.721 | −2.365 to 0.922 | 0.386 | −17.286 to 15.844 | |
| SV (mL) | AI1 vs. Manual | 8.594 | 5.581 to 11.608 | <0.001 | −21.776 to 38.964 |
| AI2 vs. Manual | −0.551 | −2.276 to 3.378 | 0.700 | −27.943 to 29.045 | |
| SI (mL/m2) | AI1 vs. Manual | 5.580 | 3.558 to 7.602 | <0.001 | −14.797 to 25.957 |
| AI2 vs. Manual | 1.359 | −0.388 to 3.105 | 0.126 | −16.245 to 18.963 | |
| CO (L/min) | AI1 vs. Manual | 0.602 | 0.370 to 0.834 | <0.001 | −1.891 to 1.955 |
| AI2 vs. Manual | 0.032 | −0.159 to 0.223 | 0.742 | −1.844 to 1.741 | |
| CI (L/min/m2) | AI1 vs. Manual | 0.353 | 0.223 to 0.483 | <0.001 | −0.956 to 1.662 |
| AI2 vs. Manual | 0.036 | −0.071 to 0.142 | 0.510 | −1.039 to 1.110 | |
| M (g) | AI1 vs. Manual | −34.159 | −39.277 to −29.040 | <0.001 | −85.746 to 17.428 |
| AI2 vs. Manual | 1.223 | −3.4313 to 5.878 | 0.603 | −45.686 to 48.132 | |
| MI (g/m2) | AI1 vs. Manual | −19.509 | −22.347 to 16.670 | <0.001 | −48.115 to 9.098 |
| AI2 vs. Manual | 0.984 | −1.487 to 3.454 | 0.432 | −23.915 to 25.882 | |
AI1 is the AI-DL simplified method; AI2 is the AI-DL standard method; Manual is the manual method. 95% CI, 95% confidence interval; AI, artificial intelligence; CI, cardiac index; CMR, cardiac magnetic resonance; CO, cardiac output; DL, deep learning; EDV, end-diastolic volume; EDVI, end-diastolic volume index; EF, ejection fraction; ESV, end-systolic volume; ESVI, end-systolic volume index; LOA, limits of agreement; LV, left ventricular; M, mass; MI, mass index; SI, stroke index; SV, stroke volume.
Discussion
Evaluation of cardiac function and structure is essential for patient monitoring, risk estimation, disease characterization and treatment assessment in cardiovascular medicine (9). The left ventricle plays a pivotal role in maintaining circulation and its parameters such as EF, SV, CO, ESV, and EDV, and mass (M) have great clinical value (14). With the rapid progression of AI, particularly DL, automatic segmentation and quantitative analysis of medical images have become feasible, efficient, and accurate, offering a powerful alternative to traditional manual approaches (9,15-17).
Clinical and imaging characteristics
The patient cohort was representative of typical clinical practice, consisting predominantly of middle-aged and elderly males, with coronary heart disease as the most common diagnosis—consistent with the known epidemiological distribution of CVD (1,2). The median resting heart rate was 75 beats/min (IQR, 66–87 beats/min), which is within an acceptable range for high-quality cine CMR acquisition without the need for pharmacological rate control.
AI-DL simplified versus standard methods
At present, CMR analysis mainly relies on manual or semi-automated software, which is labor-intensive and operator-dependent (18-20). AI refers to the use of computer technology to simulate cognitive skills similar to humans (21), whereas DL belongs to an important branch of AI, defined as the machine learning algorithm for processing neural networks, with deep structure or multiple hidden layers of neural networks, so it is also called deep neural networks (9). Automated analysis reduces operator error and time cost, allowing measurements obtained within seconds (19,22). In our study, AL-DL reduced analysis time from nearly 4 minutes manually to 50 seconds (Z=–8.855, P<0.001), consistent with previous literature reports, reflecting the high efficiency of AI (23).
The Society of Cardiovascular Magnetic Resonance (SCMR) recognizes two main methods for quantifying cardiac functional and structural parameters in CMR: the simplified method and the standard method. There is still no uniformly accepted convention on the inclusion of trabecular tissue and papillary muscles in the myocardial mass (12,13). Manual analysis based on the simplified method remains widely used (24); some can achieve the standard methods, but it is time-consuming, operator-dependent, difficult to accurately annotate, and often requires manual correction (25,26). Given the irregular morphology and variability of TPM, especially at basal and apical slices, manual delineation can be challenging (27). The simplified method attributes TPM to the blood pool, leading to overestimation of ventricular volume and underestimation of myocardial mass. In contrast, the standard method classifies TPM as myocardium, improving anatomical consistency and aligning with expert consensus (4). Our findings demonstrate that AI-DL can implement both strategies automatically with minimal correction, thus overcoming key limitations of manual approaches.
Comparison, consistency, correlation, and agreement with different methods
It is sometimes difficult to ensure the consistency and accuracy of the manual method in practice, especially when dealing with complex anatomical structures. DL has high accuracy and efficiency in image detection, segmentation, registration, and feature classification (28). Accurate segmentation of medical images is the foundation for accurate diagnosis, whereas Convolutional Neural Network (CNN) can achieve the same segmentation accuracy as expert doctors in left and right ventricular cardiac membrane tracking (27). The previous literature (4,9,19,23,27,29) has mainly focused on technical metrics such as Dice similarity coefficient (DSC), average symmetric surface distance (ASSD), and Hausdorff distance (HD), as well as clinical indicators such as LV volume, mass, and EF to evaluate the value of AI in CMR analysis. These studies have demonstrated that AI is capable of fast, accurate, robust, and reliable automatic segmentation of cardiac images. The quantitative measurements of cardiac functional structure analysis obtained using AI show good consistency and correlation with the real value or manual measurement.
Compared with manual standard analysis, the AI-DL simplified method underestimated EF and mass indices, while overestimating volumetric indices, with significant statistical differences (all P<0.001). By contrast, the AI-DL standard method showed no significant differences from manual results (all P>0.05), indicating greater similarity. As described in Results, the standard method exhibited visually better contour consistency than the simplified method.
Our data indicate that TPM have a significant impact on the quantification of LV function and structure. This observation aligns with previous work. Riffel et al. (13) found that standard analysis yielded smaller EDV and ESV but higher EF and LV-M compared with the simplified method by manual post-processing (all P<0.001): LV-EDV: 149.7±32.5 vs. 160.5±35.0 mL; LV-ESV: 48.7±14.5 vs. 56.4±16.7 mL; LV-EF: 67.7%±5.4% vs. 65.1%±5.6%; LV-M: 122.9±32.0 vs. 111.5±29.1 g. There was a strong correlation between the results for LV functional structural parameters between the two methods (all r>0.9), with all results corroborated by our AI-based validation. In this study, AI-DL confirmed the differences between the two methods. The AI-DL simplified method showed moderate to good consistency (ICC =0.690–0.928) and strong to excellent correlation (r=0.726–0.944), whereas the standard method showed higher consistency (ICC =0.798–0.978) and stronger correlation (r=0.772–0.952). Bland-Altman plots showed that most of the differences between the AI-DL (simplified method/standard method) and the manual method were within the 95% LOA, indicating good agreement. Assadi et al. (29) found in a multi-center CMR cohort study of about 814 patients that LV parameters showed an excellent correlation between AI automatic and manual analysis [left ventricular end-diastolic volume (LVEDV), r=0.98; left ventricular end-systolic volume (LVESV), r=0.94; left ventricular stroke volume (LVSV), r=0.94; LVEF, r=0.91] and the difference was small (LVEDV −3.2 mL; LVESV 1.9 mL; LVSV −5.1 mL; LVEF −1.5%). However, in a study of 600 cases conducted by Bai et al. (19), based on short-axis cardiac cine images, the mean absolute differences between AI automatic and manual analysis measurements were: LVEDV 6.1 mL, LVESV 5.3 mL, and left ventricular mass (LVM) 6.9 g, respectively. The above results show that the automatic and manual analysis of LV function and structure parameters have good correlation and consistency, and the repeatability is high, which is consistent with the results of this study. The r values (0.931, 0.927, 0.936) and difference values (0.376%, –0.630 mL, –2.710 mL) of EF, EDV, and ESV of the standard method and manual method were similar to those reported in the above literature, and the ICC and r values of the standard method parameters were higher than those of the simplified method. The difference was significantly smaller and the 95% LOA was narrower, so the difference was more easily accepted in clinical practice.
Limitations
This study had several limitations, as follows:
Single-center retrospective design. The study was conducted at a single institution, and potential selection bias cannot be fully excluded. Multi-center prospective validation is warranted.
Modest sample size. Although 104 patients represent a clinically diverse cohort, larger studies are needed to confirm generalizability across rarer pathologies.
Manual reference contours were generated by consensus of two readers, but formal inter-observer reproducibility metrics were not calculated. We have now provided post-hoc inter-observer reproducibility metrics for a subset of 20 patients; these results (ICC =0.773–0.948) confirm the stability of the manual reference standard.
Absence of manual simplified comparator. The manual analysis employed only the standard method; a matched simplified manual dataset would be required to fully validate the AI-DL simplified approach. The reported differences for the simplified method reflect systematic bias due to TPM handling convention rather than algorithmic inaccuracy.
Manual reference data were generated using two different vendor workstations (Philips IntelliSpace and GE Advanced Workstation). Although both platforms employ similar semi-automated contouring with manual correction, subtle algorithmic differences in initial contour proposals could introduce minor systematic variations in the reference standard. Given the sample size (90 Philips, 14 GE), stratified analysis by vendor was not feasible, resulting in insufficient statistical power.
Future technical improvements
Future technical improvements may further enhance the clinical utility of AI-DL in CMR. These include training on multi-vendor and multi-field-strength datasets to improve generalizability, incorporating attention mechanisms to better handle basal slice ambiguity, and extending the pipeline to include right ventricular and atrial segmentation. Additionally, integrating AI-DL analysis with emerging sequences such as 4D flow and myocardial strain imaging could provide a comprehensive cardiac assessment in a single automated workflow.
Future work may also incorporate uncertainty estimation or quality control modules to automatically identify cases requiring manual correction, thereby improving reliability in real-world clinical deployment.
Conclusions
Cine CMR represents the gold standard for the quantitative assessment of LV functional and structural parameters and is critically important in clinical practice (30). The traditional manual method is time-consuming, labor-intensive, and less reproducible. In contrast, AI-DL significantly shortens analysis time while maintaining strong correlation and consistency with manual results. Notably, the standard method achieves smaller parameter differences and narrower LOA than the simplified method, making it more clinically acceptable. Therefore, both the simplified and standard AI-DL methods can provide accurate, reliable, and efficient quantification of LV function and structure, with the standard method offering superior concordance. These findings underscore the clinical value of AI-DL in assisting CMR-based cardiac functional and structural analysis.
Supplementary
The article’s supplementary files as
Acknowledgments
None.
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 The Third Affiliated Hospital of Sun Yat-sen University Yuedong Hospital (No. II2026-004-01) and all patients provided written informed consent.
Footnotes
Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0364/rc
Funding: This study was supported by the Guangdong Medical Association Clinical Research Fund of The Third Affiliated Hospital of Sun Yat-sen University Yuedong Hospital (No. 2025YX-A1007).
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0364/coif). All authors report that this study was supported by the Guangdong Medical Association Clinical Research Fund of The Third Affiliated Hospital of Sun Yat-sen University Yuedong Hospital (No. 2025YX-A1007). The authors have no other conflicts of interest to declare.
Data Sharing Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0364/dss
References
- 1.Flora GD, Nayak MK. A Brief Review of Cardiovascular Diseases, Associated Risk Factors and Current Treatment Regimes. Curr Pharm Des 2019;25:4063-84. 10.2174/1381612825666190925163827 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Roth GA, Mensah GA, Johnson CO, Addolorato G, Ammirati E, Baddour LM, et al. Global Burden of Cardiovascular Diseases and Risk Factors, 1990-2019: Update From the GBD 2019 Study. J Am Coll Cardiol 2020;76:2982-3021. 10.1016/j.jacc.2020.11.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Dissabandara T, Lin K, Forwood M, Sun J. Validating real-time three-dimensional echocardiography against cardiac magnetic resonance, for the determination of ventricular mass, volume and ejection fraction: a meta-analysis. Clin Res Cardiol 2024;113:367-92. 10.1007/s00392-023-02204-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Schilling M, Unterberg-Buchwald C, Lotz J, Uecker M. Assessment of deep learning segmentation for real-time free-breathing cardiac magnetic resonance imaging at rest and under exercise stress. Sci Rep 2024;14:3754. 10.1038/s41598-024-54164-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Ochs A, Riffel J, Ochs MM, Arenja N, Fritz T, Galuschky C, Schuster A, Bruder O, Mahrholdt H, Giannitsis E, Frey N, Katus HA, Buss SJ, André F. Myocardial mechanics in dilated cardiomyopathy: prognostic value of left ventricular torsion and strain. J Cardiovasc Magn Reson 2021;23:136. 10.1186/s12968-021-00829-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Shoaib MA, Hossain MB, Hum YC, Chuah JH, Mohd Salim MI, Lai KW. Speckle Noise Diffusion in Knee Articular Cartilage Ultrasound Images. Curr Med Imaging 2020;16:739-51. 10.2174/1573405615666190903143330 [DOI] [PubMed] [Google Scholar]
- 7.Celebi AS, Yalcin H, Yalcin F. Current cardiac imaging techniques for detection of left ventricular mass. Cardiovasc Ultrasound 2010;8:19. 10.1186/1476-7120-8-19 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Aromiwura AA, Cavalcante JL, Kwong RY, Ghazipour A, Amini A, Bax J, Raman S, Pontone G, Kalra DK. The role of artificial intelligence in cardiovascular magnetic resonance imaging. Prog Cardiovasc Dis 2024;86:13-25. 10.1016/j.pcad.2024.06.004 [DOI] [PubMed] [Google Scholar]
- 9.Shoaib MA, Chuah JH, Ali R, Hasikin K, Khalil A, Hum YC, Tee YK, Dhanalakshmi S, Lai KW. An Overview of Deep Learning Methods for Left Ventricle Segmentation. Comput Intell Neurosci 2023;2023:4208231. 10.1155/2023/4208231 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Fadil H, Totman JJ, Hausenloy DJ, Ho HH, Joseph P, Low AF, Richards AM, Chan MY, Marchesseau S. A deep learning pipeline for automatic analysis of multi-scan cardiovascular magnetic resonance. J Cardiovasc Magn Reson 2021;23:47. 10.1186/s12968-020-00695-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Chen Z, Chen X, Liu Y, Chen EZ, Chen T, Sun S. Enhancing Cardiac MRI Segmentation via Classifier-Guided Two-Stage Network and All-Slice Information Fusion Transformer. In: Wu S, Shabestari B, Xing L, editors. Applications of Medical Artificial Intelligence: Second International Workshop, AMAI 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 8, 2023, Proceedings. Cham: Springer; 2024:145-54. [Google Scholar]
- 12.Schulz-Menger J, Bluemke DA, Bremerich J, Flamm SD, Fogel MA, Friedrich MG, Kim RJ, von Knobelsdorff-Brenkenhoff F, Kramer CM, Pennell DJ, Plein S, Nagel E. Standardized image interpretation and post-processing in cardiovascular magnetic resonance - 2020 update : Society for Cardiovascular Magnetic Resonance (SCMR): Board of Trustees Task Force on Standardized Post-Processing. J Cardiovasc Magn Reson 2020;22:19. 10.1186/s12968-020-00610-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Riffel JH, Schmucker K, Andre F, Ochs M, Hirschberg K, Schaub E, Fritz T, Mueller-Hennessen M, Giannitsis E, Katus HA, Friedrich MG. Cardiovascular magnetic resonance of cardiac morphology and function: impact of different strategies of contour drawing and indexing. Clin Res Cardiol 2019;108:411-29. 10.1007/s00392-018-1371-7 [DOI] [PubMed] [Google Scholar]
- 14.Frangi AF, Niessen WJ, Viergever MA. Three-dimensional modeling for functional analysis of cardiac images: a review. IEEE Trans Med Imaging 2001;20:2-25. 10.1109/42.906421 [DOI] [PubMed] [Google Scholar]
- 15.Liu Z, Sunar MS, Tan TS, Hitam WHW. Deep learning for retinal vessel segmentation: a systematic review of techniques and applications. Med Biol Eng Comput 2025;63:2191-208. 10.1007/s11517-025-03324-y [DOI] [PubMed] [Google Scholar]
- 16.Meng LK, Khalil A, Ahmad Nizar MH, Nisham MK, Pingguan-Murphy B, Hum YC, Mohamad Salim MI, Lai KW. Carpal Bone Segmentation Using Fully Convolutional Neural Network. Curr Med Imaging Rev 2019;15:983-9. 10.2174/1573405615666190724101600 [DOI] [PubMed] [Google Scholar]
- 17.Duan T, Chen W, Ruan M, Zhang X, Shen S, Gu W. Unsupervised deep learning-based medical image registration: a survey. Phys Med Biol 2025. doi: . 10.1088/1361-6560/ad9e69 [DOI] [PubMed] [Google Scholar]
- 18.Bernard O, Lalande A, Zotti C, Cervenansky F, Yang X, Heng PA, et al. Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved? IEEE Trans Med Imaging 2018;37:2514-25. 10.1109/TMI.2018.2837502 [DOI] [PubMed] [Google Scholar]
- 19.Bai W, Sinclair M, Tarroni G, Oktay O, Rajchl M, Vaillant G, et al. Automated cardiovascular magnetic resonance image analysis with fully convolutional networks. J Cardiovasc Magn Reson 2018;20:65. 10.1186/s12968-018-0471-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Bhuva AN, Bai W, Lau C, Davies RH, Ye Y, Bulluck H, et al. A Multicenter, Scan-Rescan, Human and Machine Learning CMR Study to Test Generalizability and Precision in Imaging Biomarker Analysis. Circ Cardiovasc Imaging 2019;12:e009214. 10.1161/CIRCIMAGING.119.009214 [DOI] [PubMed] [Google Scholar]
- 21.Amisha Malik P, Pathania M, Rathaur VK. Overview of artificial intelligence in medicine. J Family Med Prim Care 2019;8:2328-31. 10.4103/jfmpc.jfmpc_440_19 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Ambale-Venkatesh B, Lima JAC. Human-in-the-Loop Artificial Intelligence in Cardiac MRI. Radiology 2022;305:80-1. 10.1148/radiol.221132 [DOI] [PubMed] [Google Scholar]
- 23.Qin W, Wu Y, Li S, Chen Y, Yang Y, Liu X, Zheng H, Liang D, Hu Z. Automated segmentation of the left ventricle from MR cine imaging based on deep learning architecture. Biomed Phys Eng Express 2020;6:025009. 10.1088/2057-1976/ab7363 [DOI] [PubMed] [Google Scholar]
- 24.Schulz-Menger J, Bluemke DA, Bremerich J, Flamm SD, Fogel MA, Friedrich MG, Kim RJ, von Knobelsdorff-Brenkenhoff F, Kramer CM, Pennell DJ, Plein S, Nagel E. Standardized image interpretation and post processing in cardiovascular magnetic resonance: Society for Cardiovascular Magnetic Resonance (SCMR) board of trustees task force on standardized post processing. J Cardiovasc Magn Reson 2013;15:35. 10.1186/1532-429X-15-35 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Lei X, Liu H, Han Y, Cheng W, Sun J, Luo Y, Yang D, Dong Y, Chung Y, Chen Y. Reference values of cardiac ventricular structure and function by steady-state free-procession MRI at 3.0T in healthy adult Chinese volunteers. J Magn Reson Imaging. 2017;45:1684-92. 10.1002/jmri.25520 [DOI] [PubMed] [Google Scholar]
- 26.Petersen SE, Aung N, Sanghvi MM, Zemrak F, Fung K, Paiva JM, Francis JM, Khanji MY, Lukaschuk E, Lee AM, Carapella V, Kim YJ, Leeson P, Piechnik SK, Neubauer S. Reference ranges for cardiac structure and function using cardiovascular magnetic resonance (CMR) in Caucasians from the UK Biobank population cohort. J Cardiovasc Magn Reson 2017;19:18. 10.1186/s12968-017-0327-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Penso M, Moccia S, Scafuri S, Muscogiuri G, Pontone G, Pepi M, Caiani EG. Automated left and right ventricular chamber segmentation in cardiac magnetic resonance images using dense fully convolutional neural network. Comput Methods Programs Biomed 2021;204:106059. 10.1016/j.cmpb.2021.106059 [DOI] [PubMed] [Google Scholar]
- 28.Piraianu AI, Fulga A, Musat CL, Ciobotaru OR, Poalelungi DG, Stamate E, Ciobotaru O, Fulga I. Enhancing the Evidence with Algorithms: How Artificial Intelligence Is Transforming Forensic Medicine. Diagnostics (Basel) 2023;13:2992. 10.3390/diagnostics13182992 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Assadi H, Alabed S, Li R, Matthews G, Karunasaagarar K, Kasmai B, Nair S, Mehmood Z, Grafton-Clarke C, Swoboda PP, Swift AJ, Greenwood JP, Vassiliou VS, Plein S, van der Geest RJ, Garg P. Development and validation of AI-derived segmentation of four-chamber cine cardiac magnetic resonance. Eur Radiol Exp 2024;8:77. 10.1186/s41747-024-00477-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Wang Z, Wang Z, An J, Yuan Y, Pang J, He Y. Cardiac compressed sensing real-time cine for the assessment of left ventricular function: a large sample size analysis. Quant Imaging Med Surg 2024;14:8785-97. 10.21037/qims-24-980 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
The article’s supplementary files as
Data Availability Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0364/dss