RPS 605 - Artificial intelligence in cardiovascular imaging

Author Block: K. Dwivedi, M. Sharkey, S. Alabed, A. Maiter, C. S. Johns, S. Rajaram, R. Condliffe, D. Kiely, A. J. Swift; Sheffield/UK
Purpose: Pulmonary Hypertension associated with Chronic Obstructive Pulmonary Disease (PH-COPD) is a heterogenous condition, with a spectrum of predominantly emphysema and some overlapping fibrosis. All patients undergo CT, but it is not used for prognostication. The study aim is to investigate the prognostic value of an AI model that quantifies the percentage of fibrosis on baseline CT, compared to radiological assessment.
Methods or Background: PH-COPD patients with baseline CT between 2001-19 were identified from the ASPIRE registry. A validated in-house PH specific deep-learning model was run and provided percentage of fibrosis by quantifying ground glass change, ground glass with reticulation, and honeycombing. Scans were scored as none/mild/moderate/severe fibrosis by sub-specialist radiologists. Cases with mean pulmonary arterial pressure ≥ 35 mmHg were classified as severe PH-COPD, and fibrosis was grouped with a threshold of 3%. Scaled cox regression and Kaplan Meier survival analysis was performed.
Results or Findings: 157 PH-CLD patients (113 severe PH-COPD) were included. AI quantified fibrosis % was a significant predictor of mortality (HR 1.46, p<0.001) .There was a significant difference (p=0.001) in survival between patients with more and less than 3% fibrosis. One and five-year survival was 84% and 35% respectively in those with <3% fibrosis and 63% and 18% respectively in those with ≥3% fibrosis. Radiologist scored mild (HR 2.05, p=0.36) and moderate (HR 2.82, p=0.045) fibrosis was a significant predictor, but not severe fibrosis.

In severe PH-COPD, radiological scoring was not a significant predictor at any level, but AI fibrosis% was a significant predictor (HR 1.37, p<0.001).
Conclusion: CT Deep learning AI model quantified fibrosis is prognostic in predicting survival and treatment response in PH-COPD and provides additional value over radiological assessment in severe PH-COPD.
Limitations: Single registry analysis, but imaging from 21 hospitals.
Funding for this study: Research conducted during post funded by UK National Institute for Health and Care Research
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethical approval was granted by the Institutional Review Board and approved by the National Research Ethics Service (16/YH/0352).

Author Block: I. H. T. Khargi¹, M. Selles¹, N. Huber², J. Browne², B. Kietselaer², T. Leiner², M. F. Boomsma¹; ¹Zwolle/NL, ²Rochester, MN/US
Purpose: To assess the performance of a novel deep learning-based cardiac metal artifact reduction algorithm (DL-C-MAR) in a retrospective comparison with unedited conventional computed tomography angiograms (CTAs) of transcatheter aortic valve implantation (TAVI) valves and phantom experiments.
Methods or Background: DL-C-MAR was trained using multiple simulated metal implants and artifacts in 1000 CTAs. Performance of DL-C-MAR was quantitatively and qualitatively investigated in 50 TAVI patients and compared to unedited conventional CTAs. To quantitatively assess image quality, noise, contrast-to-noise ratio (CNR), artifact index (AI), and artifact volume were calculated. Diameters of the valve struts were also measured. Images were qualitatively rated on overall image quality, extent of metal artifacts and valve leaflet definition by two readers on a four-point scale. Phantom experiments were conducted using four different size steel cylinders. Diameters of the cylinders were measured by two readers and compared to their conventional counterparts and the ground truth. All images were visually screened for presence of hallucinations.
Results or Findings: In the CTAs, DL-C-MAR resulted in a higher CNR (9.1±5.8 vs. 7.9±4.8), and lower noise (57.2±33.9 vs. 82.1±54.0), AI (53.0±36.0 vs. 75.7±55.9), and artifact volume (0.02±0.12mL vs. 0.06±0.42mL) compared to unedited conventional CTAs (all p<0.001). The strut diameter also decreased after DL-C-MAR (1.52±0.26mm vs. 2.05±0.46mm, p=0.005). Initial results from the qualitative analysis suggest increased valve leaflet definition and decreased metal artifact severity after DL-C-MAR. In the phantom scans, DL-C-MAR decreased cylinder diameter by 7-67% (p<0.001), bringing them closer to the ground truth. No hallucinations were observed.
Conclusion: DL-C-MAR increases image quality and reduces metal artifacts in CTAs after TAVI implantation and does not seem to hallucinate on clinical or phantom images.
Limitations: This study did not include impact on clinical decision-making outcomes.
Funding for this study: No funding was received for this study
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study was reviewed and approved as exempt with waived informed consent. Reference no.: RPR -2024-00000086

Author Block: M. Finazzo¹, M. M. Lagana², F. Graziano³, F. Pinto², C. Duranti¹, F. Finazzo¹; ¹Palermo/IT, ²Milan/IT, ³Monza/IT
Purpose: The assessment of coronary stents using Coronary CT Angiography (CCTA) can be challenging.
Deep Learning Reconstruction (DLR) is an innovative CT image reconstruction method that reduces noise, enhancing image quality.
This study aims to evaluate whether DLR improves diagnostic confidence in coronary stent evaluation using CCTA images, compared to hybrid iterative reconstruction (HIR) and model-based iterative reconstruction (MBIR).
Methods or Background: CCTA images of 20 patients with 35 stents were evaluated retrospectively using three reconstruction methods: HIR, MBIR, and DLR.
All examinations were conducted using a 320-row whole-heart CT scanner.
The diagnostic confidence of the images obtained with each reconstruction method was evaluated using a Likert score (1=non-diagnostic, 2=poor, 3=acceptable, 4=good, 5=excellent).
Stents were divided into proximal and distal according to their location. Stents located in the proximal and intermediate segments of the coronary arteries were considered proximal; stents situated in the distal segments of the main coronary arteries and side branches were considered distal.
A cumulative liked mixed model was created in Rstudio version 4.3.1 to examine the differences across
reconstruction methods while accounting for the stent position, to explore its potential effect on
diagnostic confidence. Post-hoc comparisons were conducted, and the p-values were adjusted using the
Tukey method.
Results or Findings: The reconstruction method had a significant impact, irrespective of stent position. Likert scores were significantly higher for DLR images compared to those reconstructed using HIR and MBIR (p<0.001), with no significant difference between HIR and MBIR (p=0.957).
Conclusion: DLR provided the best diagnostic confidence and significantly enhanced the evaluation of coronary stents. As a further development, Super Resolution DLR, a new reconstruction algorithm, could improve spatial resolution, thereby increasing diagnostic confidence in coronary stents.
Limitations: Limited number of patients.
Qualitative analysis only
Funding for this study: No funding
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: It's a non-pharmacological retrospective observational study which have been approved by the local ethics committee

Author Block: P. Arnold, M. Russe, E. Kotter, M. T. Hagar; Freiburg/DE
Purpose: The Coronary Artery Disease-Reporting and Data System (CAD-RADS) 2.0 offers standardized guidelines for interpreting coronary artery disease in cardiac computed tomography (CT). Accurate and consistent CAD-RADS 2.0 scoring is crucial for comprehensive disease characterization and clinical decision-making. This study investigates the capability of large language models (LLMs) to autonomously generate CAD-RADS 2.0 scores from cardiac CT reports.
Methods or Background: A dataset of 200 synthetic cardiac CT reports was created to evaluate the performance of several state-of-the-art LLMs in generating CAD-RADS 2.0 scores via in-context learning. The tested models included GPT-3.5, GPT-4o, Mistral 7b, Mixtral 8x7b, LLama3 8b, LLama3 8b with a 64k context length, and LLama3 70b. The generated scores from each model were compared to the ground truth, which was provided by an independent committee of two board-certified cardiothoracic radiologists.
Results or Findings: The GPT-4o model and Llama3 70b achieved the highest accuracy in generating full CAD-RADS 2.0 scores including all modifiers, with a performance rate of 93% and 92.5% respectively, followed by Mixtral 8x7b with 78%. In contrast, less advanced LLMs, such as Mistral 7b and GPT-3.5 provided poor performance (16%). Llama3 8b demonstrated intermediate results, with an accuracy of 41.5%.
Conclusion: Advanced LLMs are capable of generating autonomously CAD-RADS 2.0 scores for cardiac CT reports with excellent accuracy, potentially enhancing both the efficiency and consistency of cardiac CT report evaluations. Open-source models not only deliver competitive accuracy but also present the benefit of local hosting, mitigating concerns around data privacy.
Limitations: To ensure data privacy and avoid ethical concerns, this study was conducted using synthetically generated cardiac CT reports. Even though these were deemed indistinguishable from real patient reports, further research is needed to validate LLM performance in real-world settings.
Funding for this study: Hans A. Krebs Medical Scientist Program (Uniklinikum Freiburg)
German Research Foundation (DFG) - SFB 1597 - 499552394
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: None

Author Block: Z. Zheng, X. Cao, W. Liu; Shanghai/CN
Purpose: HR-MRI provided a non-invasive and radiation-free method for assessing atherosclerosis, with strong advantages for vessel wall visualization. However, efficient segmentation and stenosis degree evaluation remained a challenging dilemma that is both labor- and time-consuming and susceptible to interobserver variability. Thus, a multi-stage deep learning architecture was developed to address above issues.
Methods or Background: The method contained three modules: artery localization, automatic segmentation, and stenosis degree evaluation modules. The 422 scans were retrospectively collected from two tertiary hospitals between 2018 and 2023 with a training-validation set (372 patients, 545 lesions) and an independent test set (50 patients, 96 lesions). An external validation set (26 patients, 42 lesions) was collected prospectively between 2023 and 2024. Subsequently, the artery segmentation and stenosis degree evaluation were compared against the ground truth, which was established by consensus among three radiologists and derived from diagnostic results obtained via DSA.
Results or Findings: The results showed outstanding performance with high DSC, IOU, and low RVE, ASSD, and HD95. The concordance correlation coefficient (CCC) was 0.985(95% CI: 0.981-0.987), 0.979(95% CI: 0.963-0.984), and 0.963(95% CI: 0.944-0.992) for volumes of artery on all datasets. Stenosis degree was evaluated on the NASCET achieved Acc of 0.8750, 0.8571, AUC of 0.89, 0.80, Sens of 0.8611, 0.9333, and Spec of 0.9167, 0.6667 on the independent test and external validation sets, respectively.
Conclusion: The method achieved no less accuracy than manual segmentation by physicians and maintained a high consistency with the DSA diagnostic criteria. In addition, by shortening diagnostic time and minimizing inter-observer variability, it offered an efficient intelligent aid in clinical practice.
Limitations: The method performed in multi-stage may take up a large amount of computational resources and modifications to the architecture are required to optimize the inference speed.
Funding for this study: This work has received funding from the National Natural Science Foundation of China (82402393, 82102132, 8237071280), the Science and Technology Commission of Shanghai Municipality (20S31904300, 22TS1400900, 23S31904100, 22ZR1409500) and the Greater Bay Area Institute of Precision Medicine (Guangzhou) (KCH2310094).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: All patients or their guardians gave informed consent to use their anonymized MRI images and clinical data for research purposes. Since all data were obtained in the course of daily work, the Ethics Committee waived the need for informed consent.

Author Block: B. Durand, V. De Villedon De Naide, T. Génisson, M. Stuber, A. Bustin, H. Cochet; Bordeaux/FR
Purpose: develop and test an AI-driven deep learning model for joint segmentation of healthy myocardium, scar tissue, and microvascular obstruction (MVO) in cardiac MRI using bright-blood phase-sensitive inversion recovery (PSIR) imaging.
Methods or Background: Current methods for scar and MVO quantification in PSIR imaging are manual or semi-automated, time-consuming, and prone to errors and variability. Using a nnUNET architecture, the model was trained on 50 PSIR exams with suspected ischemic heart disease and evaluated on a test set of 20 cases. Data augmentations were applied, and manual segmentations by radiologists were used for comparison. To maximize performance, a joint segmentation approach was employed, and both magnitude and phase maps were used together.
Results or Findings: The AI model demonstrated excellent performance despite only 50 exams in training, in segmenting healthy myocardium (median Dice score 0.96) and good results for scar segmentation (median Dice score 0.75). MVO detection was successful in 2 out of 3 cases. Inference time was under 5 seconds per exam, and no false positives were identified outside the myocardium,
Conclusion: AI-driven approach showed robust segmentation of myocardium and scar tissue, with promising results in MVO detection. It could streamline clinical workflows for myocardial infarction assessment by reducing time-counsuming manual segmentations.
Limitations: The model was trained on a small dataset from post-ischemic patients, which limits its generalizability to other cardiac conditions such as hypertrophic cardiomyopathy or infiltrative diseases, where scarring patterns differ. We plan to expand the training population to improve the model's performance in non-ischemic cardiomyopathy.
Funding for this study: This research was supported by funding from the French National Research Agency under grant agreement ANR-22-CPJ2-0009-01, and from the European Research Council (ERC) grant "SMHEART" under the European Union’s Horizon 2020 research and innovation programme (grant agreement No101076351).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Biomedical Research Ethics Committee and all participants provided informed consent for participation.