RPS 1303 - The evolving impact of artificial intelligence (AI) in cardiac imaging

Author Block: S. Gowda, V. Raj, R. Kothari; Bengaluru/IN
Purpose: Accurate donor heart size measurement is crucial for successful heart transplantation (HT). Traditional weight-based donor-to-recipient (D-R) size matching in paediatric HT has poor correlation with cardiac size and significantly restricts the donor pool. The aim of the study is to develop a novel predictive model to accurately calculate Total Cardiac Volume (TCV) tailored to the Indian population, aiming to expand the donor pool and reduce size mismatches.
Methods or Background: This multi-centre study incorporated paediatric and young adults (ages 0-30) with normal CT chest angiograms. TCV was predicted using common variables such as weight, height, gender and cardiac width on chest radiograph (CXR) with CT derived TCV (3D segmentation) as the gold standard. Three predictive models were analysed, and subjects were split into training and testing data.
Model A- weight only
Model B- weight, height, gender and age
Model C- Model B plus horizontal cardiac width from CXR.
Results or Findings: Model C showed highest accuracy in predicting TCV with an R² of 0.94 for training data and 0.91 for testing data, with mean absolute percentage error (MAPE) of 3%. Model A was weakest with an R² of 0.82 for training data, 0.68 for testing data, and a MAPE of 6.3%.
Conclusion: TCV can be accurately predicted using readily available donor metrics. The proposed D-R TCV matching model can significantly expand the donor pool and improve size matching in paediatric heart transplantation in India.
Limitations: Single centre study, which may also have an in-built case selection bias.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Owing to the retrospective nature of the study, ethical committee approval was waived off by the institutional ethics committee.

Author Block: C. G. Glessgen¹, L. A. Crowe¹, J. Wetzl², M. Schmidt², S. S. Yoon², J-P. Vallee¹, J-F. Deux¹; ¹Geneva/CH, ²Erlangen/DE
Purpose: Cardiac MRI (CMR) is demanding due to the number of planning steps and parameters requiring continuous monitoring. The high mistake risk can impact procedure quality, scan times, and data homogeneity. The impact of an AI-based automated CMR planning software on procedure errors and scan times compared to human-guided examinations is evaluated.
Methods or Background: Consecutive patients undergoing non-stress CMR were prospectively enrolled into two acquisition modes: manual or automated utilizing prototype software (Siemens Healthineers, Erlangen, Germany). Patients with pacemakers or targeted indications were excluded. All underwent the same CMR protocol with contrast administration, in breath-hold (BH) or free breathing (FB). Supervising radiologists recorded procedure errors (plane prescription, forgotten views, incorrect propagation of a cardiac plane, field-of-view mismanagement). Scan times and Dead Phase (non-acquisition portion) were computed from scanner logs. Most data were non-normally distributed and compared using nonparametric tests.
Results or Findings: Eighty-two patients (mean age, 51.6 years; 56 male) were included. Forty-four patients underwent automated CMR and 38 manual CMR. The rate of procedure errors per CMR was lower (p=0.01) in automated (0.45) than in manual (1.13). The ratio of error-free examinations was higher (p=0.03) in automated (31/44; 70.5%) than in manual (17/38; 44.7%). Automated studies were shorter than manual studies in FB (30.3 vs. 36.5 minutes, p<.001) but had similar durations in BH (42.0 vs. 43.5 minutes, p=0.42). Dead Phase was lower in automated studies for both FB and BH strategies (p<.001).
Conclusion: AI-based automation performed cardiac MRI studies at a clinical level with fewer planning errors and improved efficiency compared to human planning.
Limitations: No reproductibility analysis of plane adjustments between radiologists was performed. Radiologists could not realistically be blinded to the study arm as automatic acquisitions were performed with almost no visible human interaction.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: All patients gave informed consent

Author Block: E. Zsarnóczay¹, A. Varga-Szemes², U. J. Schoepf², S. Rapaka³, N. Fink², M. Vecsey-Nagy², P. Sharma³, P. Maurovich-Horvat¹, T. S. Emrich²; ¹Budapest/HU, ²Charleston, SC/US, ³Princeton, NJ/US
Purpose: To determine whether artificial intelligence (AI)–based fully automated assessment of left atrioventricular coupling index (LACI) can provide incremental value above other traditional risk factors for predicting mortality among patients with severe aortic stenosis (AS)
undergoing coronary CT angiography (CCTA) before transcatheter aortic valve replacement (TAVR).
Methods or Background: This retrospective study evaluated patients with severe AS who underwent CCTA examination before TAVR between 2014 and 2019. An AI-prototype software fully automatically calculated left atrial (LA) and left ventricular (LV) end-diastolic volumes and LACI was defined as the ratio between them. Clinical parameters, the Society of Thoracic Surgeons Predicted Risk of Mortality (STS-PROM) risk score, and all-cause mortality after TAVR were recorded. Uni- and multivariate Cox proportional hazard methods were used to identify the predictors of mortality in models adjusting for relevant significant parameters, STS-PROM score, and patients with preserved LV ejection fraction (EF).
Results or Findings: A total of 656 patients (77 years [IQR, 71-84 years]; 387 [59.0%] male) were included. The all-cause mortality rate was 21.6% over a median follow-up time of 24 (10–40) months. When adjusting for clinical confounders, LACI≥43.7% was found to independently predict mortality (adjusted HR, 1.52, [95CI: 1.03,2.22]; p=0.032). After adjusting for the STS-PROM score in a separate model, LACI ≥43.7% remained an independent prognostic parameter (adjusted HR, 1.47, [95CI: 1.03,2.08]; p=0.031). In a sub-analysis of patients with preserved LVEF, LACI remained a significant predictor (adjusted HR, 1.72 [95CI: 1.02,2.89]; p=0.042).
Conclusion: AI-based fully automated assessment of LACI can be used independently to predict mortality in patients undergoing TAVR, including those with preserved LVEF.
Limitations: This study was performed in a single-center and single-vendor setting, using an AI-powered software prototype, selection bias may exist because only patients with available outcomes data were included.
Funding for this study: Not applicable.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Not applicable.

Author Block: L. R. M. Lanzafame¹, T. D'Angelo¹, A. Othman², C. Booz³; ¹Messina/IT, ²Mainz/DE, ³Frankfurt/DE
Purpose: This study aimed to evaluate the impact of a deep learning-based denoising (DLD) technique on image quality and diagnostic accuracy for the assessment of coronary arteries in pre-procedural transcatheter aortic valve implantation (TAVI) CT planning.
Methods or Background: A retrospective analysis was conducted on 200 patients with severe aortic stenosis who underwent CT scans for TAVI planning between October 2022 and April 2024. Conventional images were reconstructed, and denoised images were generated using DLD model. Objective image quality was assessed by measuring the mean Hounsfield unit (HU) and standard deviation (SD) in the aortic root, coronary arteries, and subcutaneous fat to calculate noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR). Two independent readers subjectively evaluated sharpness, noise, vascular contrast, and overall image quality using a 5-point Likert scale. Diagnostic performance was compared between original and denoised images by assessing accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), using invasive coronary angiography as the reference standard.
Results or Findings: Denoised images demonstrated significantly improved SNR (37.5 ± 12.8 vs. 12.3 ± 4.1) and CNR (45.3 ± 15.4 vs. 14.7 ± 4.4), along with reduced noise levels (16.9 ± 7.9 vs. 47.9 ± 11.6 HU) (all p<0.001). Subjective evaluations also favored denoised images in terms of sharpness, noise reduction, contrast, and overall quality (all p<0.001). DLD reconstructions revealed higher diagnostic performance, showing a on a per-segment basis sensitivity of 95.9%, specificity of 94.3%, PPV of 86.5%, NPV of 98.4%, and accuracy of 94.8%.
Conclusion: The DLD algorithm significantly improves image quality and diagnostic accuracy in pre-TAVI CT imaging for coronary artery evaluation.
Limitations: The retrospective design prevented evaluation of image quality at reduced radiation doses. Furthermore, the results are specific to our acquisition protocol.
Funding for this study: This research did not receive external funding.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Ethics Committee of Johannes Gutenberg University of Mainz (Ref. Nr. 2022-16477_1)

Author Block: M. Manzke¹, F. C. Laqua², B. Böttcher¹, A-C. Klemenz¹, M-A. Weber¹, B. Baeßler², F. G. Meinel¹; ¹Rostock/DE, ²Würzburg/DE
Purpose: To investigate the reproducibility of radiomic features in myocardial native T1 and T2 mapping.
Methods or Background: Cardiac MRI T1 maps from 50 healthy volunteers (29 women and 21 men, mean age 39.4 ± 13.7 years) underwent two identical cardiac MRI examinations at 1.5T. The protocol included native T1 and T2 mapping in both short-axis and long-axis orientation. For T1 mapping, we investigated standard (1.9 x 1.9 mm) and high (1.4 x 1.4 mm) spatial resolution. After manual segmentation of the left ventricular myocardium, 100 radiomic features from seven feature classes were extracted and analyzed. Test–retest repeatability of radiomic features was assessed using the intraclass correlation coefficient (ICC) and classified as poor (ICC <0.50), moderate (0.50–0.75), good (0.75–0.90) and excellent (>0.90).
Results or Findings: For T1 maps acquired in short-axis orientation at standard resolution, repeatability was excellent for 6 features, good for 29 features, moderate for 19 features and poor for 46 features. We identified 15 features from 6 classes which showed good to excellent reproducibility for T1 mapping in all resolutions and all orientations. For short-axis T2 maps, repeatability was excellent for 6 features, good for 25 features, moderate for 23 features and poor for 46 features. 12 features from 5 classes were found to have good to excellent repeatability in T2 mapping independent of slice orientation.
Conclusion: We have identified a subset of radiomic features with good to excellent repeatability independent of slice orientation and spatial resolution. We recommend using these features for further radiomics research in myocardial T1 and T2 mapping.
Limitations: This study was limited to healthy volunteers. The reproducibility of radiomic features in patients with diffuse or focal myocardial disease cannot be directly concluded.
Funding for this study: The study was in part funded by the Federal Ministry of Education and Research (BMBF) through the Network University Medicine „NUM 2.0“ (grant number 01KX2121).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the institutional review board and written informed consent was obtained from all volunteers prior to enrollment.

Author Block: B. Böttcher¹, K. K. Deyerberg¹, A-C. Klemenz¹, L-M. Watzke¹, M. Gorodezky², M. Manzke¹, M-A. Weber¹, F. G. Meinel¹; ¹Rostock/DE, ²Munich/DE
Purpose: Plane positioning in cardiac magnetic resonance imaging (cMRI) is crucial for diagnostic image quality and comparability of cardiac functional parameters in follow-up exams. Manual planning is influenced by user’s training making it susceptible for inter-reader variability and errors. This prospective cohort study aims to investigate the reproducibility of an artificial intelligence-based planning approach against state-of-the-art manual plane prescription.
Methods or Background: 25 healthy participants (mean age 41.5, range: 23-65 years, mean BMI 25.2 kg/m²) underwent two identical cMRI exams on a 1.5T scanner (Signa Artist, GE HealthCare). Short axis, 2-, 3- and 4-chamber planes (FOV: 34x34cm2, matrix size: 200x224, slice thickness: 8mm) were acquired using an AI-based planning tool (TeslaFlow prototype, GE HealthCare) and manual planning. Short axis left ventricular volumetric analysis (end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV) and ejection fraction (EF)) were performed using an established post-processing software (cvi42, Circle Cardiovascular Imaging). The Wilcoxon matched-pairs signed rank test with a significance level of p≤0.05 was used to compare the first to the second exam for both the manual and automated planning.
Results or Findings: Volumetric parameters calculated on manual and AI-assisted planned images showed following median of differences between both scans: EDV -5.0ml (p=0.220), -2.8ml (p=0.474); ESV 0.1ml (p=0.560), 2.0ml (p=0.367); SV -3.6ml (p=0.096), -4.2ml (p=0.043) and EF -1.2% (p=0.329), -2.8% (p=0.045), respectively. The only statistically significant differences were observed in SV and EF for AI-based planning, though the deviation is not clinically relevant.
Conclusion: AI-based planning for cMRI showed high reproducibility without clinically relevant variability between follow-up scans. This novel technique can simplify and accelerate cMRI maintaining high diagnostic quality.
Limitations: This study was conducted on a cohort of healthy individuals at a single MRI scanner, provided by a single vendor.
Funding for this study: None.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was designed as a prospective, single-center cohort study and approved by the responsible institutional review board of the Medical University Center of Rostock.

Author Block: F. Fanelli, D. De Santis, L. Pugliese, G. G. Bona, C. Santangeli, T. Polidori, G. Tremamunno, D. Caruso, A. Laghi; Rome/IT
Purpose: Cardiovascular magnetic resonance (CMR) is the reference standard for the assessment of cardiac function, achieved through conventional balanced steady-state free precession (bSSFP) cine sequences, which represent a considerable part of the CMR exams, contributing to patient discomfort. The aim of our study was to evaluate diagnostic accuracy and image quality of deep-learning(DL)cine sequences for LV and RV parameters compared to bSSFP cine sequences in CMR.
Methods or Background: From January to April 2024, patients with clinically indicated CMR were prospectively included. LV and RV were segmented from short-axis bSSFP and DL cine sequences. LV and RV end-diastolic volume, end-systolic volume, stroke volume, ejection fraction, and LV end-diastolic mass were calculated. The acquisition time of both sequences was registered. Results were compared with paired-samples t-test or Wilcoxon signed-rank test. Agreement between DL cine and bSSFP was assessed using Bland-Altman plots. Image quality was graded by two readers based on blood-to-myocardium contrast, endocardial edge definition, and motion artifacts, using a 5-point Likert scale (1= insufficient quality; 5= excellent quality).
Results or Findings: Sixty-two patients were included (mean age: 47±17 years, 41 men). No significant differences between DL cine and bSSFP were found for all LV and RV parameters (P≥ .176). DL cine was significantly faster (1.35 ±.55 m vs 2.83 ± .79 m; P< .001). The agreement between DL cine and bSSFP was strong, with near-zero bias and good limits of agreement. Overall image quality was comparable (median: 5, IQR: 4-5; P= .330), while endocardial edge definition of DL cine (median: 4, IQR: 4-5) was lower than bSSFP (median: 5, IQR: 4-5; P= .002).
Conclusion: DL cine allows fast and accurate quantification of LV and RV parameters and comparable image quality with conventional bSSFP.
Limitations: Not applicable
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study has been approved by local Ethics committee .

Author Block: N. Tomizawa, Y. Nozaki, R. Fan, Y. Kawaguchi,, K. Takamura, F. Shinichiro, K. Kumamaru, T. Minamino, S. Aoki; Bunkyo-Ku/JP
Purpose: The purpose of this study was to compare the diagnostic performance of CT fractional flow reserve (CT-FFR) using model-based iterative reconstruction (MBIR) and super-resolution deep learning reconstruction (SR-DLR) to detect functionally significant stenosis as assessed by invasive FFR.
Methods or Background: This single-center retrospective study included 79 patients (mean age, 70 years ± 11 [SD]; 57 men) who underwent coronary CT angiography showing intermediate stenosis (30%‒70% stenosis) and subsequent invasive FFR between February 2022 and March 2024. Vessels with heavy calcification were not excluded from the analysis. Computational fluid dynamics was used to calculate the CT-FFR using MBIR and SR-DLR images. Per-vessel diagnostic performance to detect FFR ≤0.80 in coronary angiography was compared by analyzing receiver operating characteristic (ROC) curves.
Results or Findings: Of the 98 vessels evaluated, 46 vessels (47%) had functionally significant stenosis. The median (interquartile range) calcium score was 462 (134–932). CT-FFR values calculated using both MBIR (mean difference: −0.088; 95% CI: −0.129, −0.048; p <0.001) and SR-DLR (mean difference: −0.026; 95% CI: −0.050, −0.002; p = 0.03) were underestimated compared to invasive FFR. The area under the ROC curve to diagnose functionally significant stenosis was higher for SR-DLR (0.88; 95% CI: 0.80, 0.95) than for MBIR (0.76; 95% CI: 0.67, 0.86; p = 0.003). CT-FFR calculated using SR-DLR had improved diagnostic accuracy (88% vs. 70%, p <0.001) and specificity (87% vs. 63%, p <0.001) over MBIR, but had similar sensitivity (89% vs. 78%, p = 0.06).
Conclusion: SR-DLR images improved the diagnostic performance of CT-FFR over MBIR images in detecting functionally significant stenosis as assessed by invasive FFR.
Limitations: This study is retrospective and used a single CT vendor. Multi-vendor multi-center study is necessary to confirm the findings.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved by the Ethics Committee of Juntendo University on May 2, 2024 (No. E23-0040-H02)

Author Block: V. Vingiani, B. Proner, N. Cortellini, R. Valletta, T. Gorgatti, A. Posteraro, V. Corato, M. Bonatti; Bolzano/IT
Purpose: This study assessed the feasibility of a customized GPT-4 model in categorising cardiac radiological reports using the Coronary Artery Disease Reporting and Data System (CAD-RADS) classification.
Methods or Background: A customised GPT-4 model was developed using the CAD-RADS 2.0-2022 guidelines, provided as a PDF, and fine-tuned on 30 clinical scenarios. The model was tested on 118 anonymized Coronary CT Angiography (CCTA) reports. Data included patient metrics and report details (e.g., length, conclusions). The reports were also reviewed by a radiologist with 9 years of experience, who categorised them according to the CAD-RADS classification. The time required for manual assessment was recorded. The GPT-4 model's performance was evaluated using Cohen's kappa for agreement and the Wilcoxon test for comparing processing times between the model and the radiologist.
Results or Findings: The mean patient age was 59.6 years (±10.9), with 41% women. Reports were authored by 10 radiologists, with 88% in Italian and 12% in German. The median report length was 1,456 characters, with conclusions in 76% of reports. The GPT-4 model showed substantial agreement with the radiologist, achieving a Cohen's kappa of 0.79 (95% CI: 0.68 - 0.89). It significantly reduced processing time, averaging 16 seconds per report compared to 57 seconds for the radiologist (P < 0.0001).
Conclusion: These findings suggest that a customized GPT-4 model is a promising tool for autonomously categorising radiological findings using the CAD-RADS classification when the original report does not include it, thereby offering a time-efficient alternative. Implementing such a system could assist clinicians and cardiologists in consistently interpreting reports by providing CAD-RADS classifications when they are not explicitly reported.
Limitations: The GPT model was fine-tuned using only 30 scenarios. A larger dataset could be used to improve the model's performance.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study was conducted in accordance with the principles of the Declaration of Helsinki

Author Block: F. Tatsugami¹, A. Streiff², T. Higaki¹, A. Labani², W. Fukumoto¹, S. El Ghannudi², K. Haioun³, K. Awai¹, M. Ohana²; ¹Hiroshima/JP, ²Strasbourg/FR, ³Tokyo/JP
Purpose: DLR-based Motion Correction (MC-DLR) for Coronary CTA has the potential to reduce/eliminate kinetic artifacts in CCTA more effectively than traditional algorithms, but its clinical impact is still unknown.
We aim to evaluate the effect of MC-DLR on coronary luminal and stenosis assessment in a varied CCTA cohort.
Methods or Background: Sixty CCTA (20 with HR<60bpm, 20 with HR=60-75, 20 with HR>75) with various degrees of stenosis (50% CAD-RADS 1 & 2, 50% CAD-RADS 3 & 4) were retrospectively selected from 2 tertiary centers.
All scans were acquired on 4th/5th-gen wide-area detector CT within 1 heartbeat.
Best phase for each included CCTA was reconstructed without and with MC-DLR, using Super Resolution DLR with 1024 matrix-size. MC-DLR subdivides the data required for volume reconstruction into smaller time sections to estimate coronary artery motion.
Four radiologists with varying levels of expertise independently and randomly reviewed all 120 datasets to:
(A) grade the luminal/wall image quality using a 3-level scale, for the 9 coronary artery segments, and
(B) assess CAD-RADS.
Statistical analysis used descriptive and Bayesian approaches.
Results or Findings: For each reader and for the pooled analysis, overall luminal/wall image quality score was significantly better with MC-DLR than without (p<0.05).
Per segment, the positive effect was more consistent on the RCA (improvement in 62% of cases) than on the LAD (44%) and the Cx (21%).
Effect of MC-DLR was non-existent in cases with absent/minimal coronary kinetic artifacts.
Non-significant changes in intra/inter-reader variability were noted in CAD-RADS 3/4.
Conclusion: MC-DLR significantly enhances coronary artery luminal and wall image quality in cases with moderate or severe kinetic artifacts, suggesting a potential clinical role in refining stenosis assessment when above >50%.
Limitations: Quantitative analysis with attenuation profile curves was not performed.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB from Strasbourg University Hospital

Author Block: Y. J. Suh¹, C. Kim², W-S. Yoo¹, J. Y. Kim³, S. Chang¹, C. H. Park⁴, S. J. Hong⁵, D. H. Yang¹, H. S. Yong¹; ¹Seoul/KR, ²Ansan/KR, ³Daegu/KR, ⁴Cheonan/KR, ⁵Guri/KR
Purpose: Coronary artery calcium (CAC) density on electrocardiogram (ECG)-gated CT has been suggested as an inverse prognostic marker for prediction of future adverse cardiovascular events. We aimed to evaluate the performance of artificial intelligence (AI)-based automated CAC density quantification on ECG-gated calcium scoring CT (CSCT) and non-ECG-gated low-dose chest CT (LDCT), using multi-institutional datasets.
Methods or Background: A total of 1,540 pairs of CSCT-LDCT scans from a multicenter database were retrospectively included. AI-based automated CAC quantification was conducted on the CSCT and LDCT. For cases with CAC score>0, mean and peak CAC density was calculated from the labeled CAC. For peak CAC density factors, a value of 1 to 4 was assigned based on the measured peak density attenuation (1: 130-199HU; 2: 200-299HU; 3: 300-399HU; 4: >400HU). The reliability and agreement of the mean CAC density and peak CAC density categories obtained from the automated scoring were analyzed compared to manual measurement, using the intraclass correlation coefficient (ICC), Bland-Altman analysis, and weighted kappa (κ) statistics, respectively.
Results or Findings: A total of 808 CSCT scans and 579 LDCT scans were positive for CAC. Automated mean density measurement demonstrated excellent ICCs on CSCT and LDCT (0.988 [95% CI, 0.986-0.989] vs. 0.956 [95% CI, 0.948-0.962]). Mean bias with 95% limits of agreement for the mean density was 0.7 ± 22.4 on CSCT and 0.6 ± 27.4 on LDCT. In terms of peak density category, automated measurement on CSCT and LDCT exhibited excellent reliability with manual measurement (weighted κ 0.980 [95% CI, 0.969-0.991) and 0.964 [95% CI, 0.945-0.982]).
Conclusion: AI-based automated CAC quantification can provide accurate and reliable measurement of CAC density on CSCT and LDCT across multi-institutional datasets.
Limitations: Prognostic value of AI-based CAC density should be further investigated.
Funding for this study: The Researcher Supporting Program funded by Korean Society of Cardiovascular Imaging (KOSCI) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT)(No. 2021R1A2C4002195)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approval numbers: KC22RIDI0156, 2021-12-027, 2022GR0064, 2021AS0371, 2022-01-001, 2021-0303, 2021-12-029, and 4-2021-1589