RESUMO
Background: Previous studies evaluated the ability of large language models (LLMs) in medical disciplines; however, few have focused on image analysis, and none specifically on cardiovascular imaging or nuclear cardiology. Objectives: This study assesses four LLMs - GPT-4, GPT-4 Turbo, GPT-4omni (GPT-4o) (Open AI), and Gemini (Google Inc.) - in responding to questions from the 2023 American Society of Nuclear Cardiology Board Preparation Exam, reflecting the scope of the Certification Board of Nuclear Cardiology (CBNC) examination. Methods: We used 168 questions: 141 text-only and 27 image-based, categorized into four sections mirroring the CBNC exam. Each LLM was presented with the same standardized prompt and applied to each section 30 times to account for stochasticity. Performance over six weeks was assessed for all models except GPT-4o. McNemar's test compared correct response proportions. Results: GPT-4, Gemini, GPT4-Turbo, and GPT-4o correctly answered median percentiles of 56.8% (95% confidence interval 55.4% - 58.0%), 40.5% (39.9% - 42.9%), 60.7% (59.9% - 61.3%) and 63.1% (62.5 - 64.3%) of questions, respectively. GPT4o significantly outperformed other models (p=0.007 vs. GPT-4Turbo, p<0.001 vs. GPT-4 and Gemini). GPT-4o excelled on text-only questions compared to GPT-4, Gemini, and GPT-4 Turbo (p<0.001, p<0.001, and p=0.001), while Gemini performed worse on image-based questions (p<0.001 for all). Conclusion: GPT-4o demonstrated superior performance among the four LLMs, achieving scores likely within or just outside the range required to pass a test akin to the CBNC examination. Although improvements in medical image interpretation are needed, GPT-4o shows potential to support physicians in answering text-based clinical questions.
RESUMO
Background: While low-dose computed tomography scans are traditionally used for attenuation correction in hybrid myocardial perfusion imaging (MPI), they also contain additional anatomic and pathologic information not utilized in clinical assessment. We seek to uncover the full potential of these scans utilizing a holistic artificial intelligence (AI)-driven image framework for image assessment. Methods: Patients with SPECT/CT MPI from 4 REFINE SPECT registry sites were studied. A multi-structure model segmented 33 structures and quantified 15 radiomics features for each on CT attenuation correction (CTAC) scans. Coronary artery calcium and epicardial adipose tissue scores were obtained from separate deep-learning models. Normal standard quantitative MPI features were derived by clinical software. Extreme Gradient Boosting derived all-cause mortality risk scores from SPECT, CT, stress test, and clinical features utilizing a 10-fold cross-validation regimen to separate training from testing data. The performance of the models for the prediction of all-cause mortality was evaluated using area under the receiver-operating characteristic curves (AUCs). Results: Of 10,480 patients, 5,745 (54.8%) were male, and median age was 65 (interquartile range [IQR] 57-73) years. During the median follow-up of 2.9 years (1.6-4.0), 651 (6.2%) patients died. The AUC for mortality prediction of the model (combining CTAC, MPI, and clinical data) was 0.80 (95% confidence interval [0.74-0.87]), which was higher than that of an AI CTAC model (0.78 [0.71-0.85]), and AI hybrid model (0.79 [0.72-0.86]) incorporating CTAC and MPI data (p<0.001 for all). Conclusion: In patients with normal perfusion, the comprehensive model (0.76 [0.65-0.86]) had significantly better performance than the AI CTAC (0.72 [0.61-0.83]) and AI hybrid (0.73 [0.62-0.84]) models (p<0.001, for all).CTAC significantly enhances AI risk stratification with MPI SPECT/CT beyond its primary role - attenuation correction. A comprehensive multimodality approach can significantly improve mortality prediction compared to MPI information alone in patients undergoing cardiac SPECT/CT.
RESUMO
Background: Non-contrast CT scans are not used for evaluating left ventricle myocardial mass (LV mass), which is typically evaluated with contrast CT or cardiovascular magnetic resonance imaging (MRI). We assessed the feasibility of LV mass estimation from standard, ECG-gated, non-contrast CT using an artificial intelligence (AI) approach and compare it with coronary CT angiography (CTA) and cardiac MRI. Methods: We enrolled consecutive patients who underwent coronary CTA, which included non-contrast CT calcium scanning and contrast CTA, and cardiac MRI. The median interval between coronary CTA and MRI was 22 days (IQR: 3-76). We utilized an nn-Unet AI model that automatically segmented non-contrast CT structures. AI measurement of LV mass was compared to contrast CTA and MRI. Results: A total of 316 patients (Age: 57.1±16.7, 56% male) were included. The AI segmentation took on average 22 seconds per case. An excellent correlation was observed between AI and contrast CTA LV mass measures (r=0.84, p<0.001), with no significant differences (136.5±55.3 vs. 139.6±56.9 g, p=0.133). Bland-Altman analysis showed minimal bias of 2.9. When compared to MRI, measured LV mass was higher with AI (136.5±55.3 vs. 127.1±53.1 g, p<0.001). There was an excellent correlation between AI and MRI (r=0.85, p<0.001), with a small bias (-9.4). There were no statistical differences between the correlations of LV mass between contrast CTA and MRI, or AI and MRI. Conclusions: The AI-based automated estimation of LV mass from non-contrast CT demonstrated excellent correlations and minimal biases when compared to contrast CTA and MRI.
RESUMO
We investigated the prognostic utility of visually estimated coronary artery calcification (VECAC) from low dose computed tomography attenuation correction (CTAC) scans obtained during SPECT/CT myocardial perfusion imaging (MPI), and assessed how it compares to coronary artery calcifications (CAC) quantified by calcium score on CTACs (QCAC). From the REFINE SPECT Registry 4,236 patients without prior coronary stenting with SPECT/CT performed at a single center were included (age: 64 ± 12 years, 47% female). VECAC in each coronary artery (left main, left anterior descending, circumflex, and right) were scored separately as 0 (absent), 1 (mild), 2 (moderate), or 3 (severe), yielding a possible score of 0-12 for each patient (overall VECAC grade zero:0, mild:1-2, moderate: 3-5, severe: >5). CAC scoring of CTACs was performed at the REFINE SPECT core lab with dedicated software. VECAC was correlated with categorized QCAC (zero: 0, mild: 1-99, moderate: 100-399, severe: ≥400). A high degree of correlation was observed between VECAC and QCAC, with 73% of VECACs in the same category as QCAC and 98% within one category. There was substantial agreement between VECAC and QCAC (weighted kappa: 0.78 with 95% confidence interval: 0.76-0.79, p < 0.001). During a median follow-up of 25 months, 372 patients (9%) experienced major adverse cardiovascular events (MACE). In survival analysis, both VECAC and QCAC were associated with MACE. The area under the receiver operating characteristic curve for 2-year-MACE was similar for VECAC when compared to QCAC (0.694 versus 0.691, p = 0.70). In conclusion, visual assessment of CAC on low-dose CTAC scans provides good estimation of QCAC in patients undergoing SPECT/CT MPI. Visually assessed CAC has similar prognostic value for MACE in comparison to QCAC.
Assuntos
Calcinose , Doença da Artéria Coronariana , Imagem de Perfusão do Miocárdio , Humanos , Feminino , Pessoa de Meia-Idade , Idoso , Masculino , Imagem de Perfusão do Miocárdio/métodos , Prognóstico , Valor Preditivo dos Testes , Tomografia Computadorizada de Emissão de Fóton Único/métodos , Doença da Artéria Coronariana/diagnóstico por imagem , Tomografia Computadorizada por Raios X/métodosRESUMO
We aimed to develop a novel deep-learning based method for automatic coronary artery calcium (CAC) quantification in low-dose ungated computed tomography attenuation correction maps (CTAC). In this study, we used convolutional long-short -term memory deep neural network (conv-LSTM) to automatically derive coronary artery calcium score (CAC) from both standard CAC scans and low-dose ungated scans (CT-attenuation correction maps). We trained convLSTM to segment CAC using 9543 scans. A U-Net model was trained as a reference method. Both models were validated in the OrCaCs dataset (n=32) and in the held-out cohort (n=507) without prior coronary interventions who had CTAC standard CAC scan acquired contemporarily. Cohen's kappa coefficients and concordance matrices were used to assess agreement in four CAC score categories (very low: <10, low:10-100; moderate:101-400 and high >400). The median time to derive results on a central processing unit (CPU) was significantly shorter for the conv-LSTM model- 6.18s (inter quartile range [IQR]: 5.99, 6.3) than for UNet (10.1s, IQR: 9.82, 15.9s, p<0.0001). The memory consumption during training was much lower for our model (13.11Gb) in comparison with UNet (22.31 Gb). Conv-LSTM performed comparably to UNet in terms of agreement with expert annotations, but with significantly shorter inference times and lower memory consumption.
RESUMO
INTRODUCTION: The degree of stenosis on coronary CT angiography (CCTA) guides referral for CT-derived flow reserve (FFRct). We sought to assess whether semiquantitative assessment of high-risk plaque (HRP) features on CCTA improves selection of studies for FFRct over stenosis assessment alone. METHODS: Per-vessel FFRct was computed in 1,395 vessels of 836 patients undergoing CCTA with 25-99% maximal stenosis. By consensus analysis, stenosis severity was graded as 25-49%, 50-69%, 70-89%, and 90-99%. HRPs including low attenuation plaque (LAP), positive remodeling (PR), and spotty calcification (SC) were assessed in lesions with maximal stenosis. Lesion FFRct was measured distal to the lesion with maximal stenosis, and FFRct<0.80 was defined as abnormal. Association of HRP and abnormal lesion FFRct was evaluated by univariable and multivariable logistic regression models. RESULTS: The frequency of abnormal lesion FFRct increased with increase of stenosis severity across each stenosis category (25-49%:6%; 50-69%:30%; 70-89%:54%; 90-99%:91%, p â< â0.001). Univariable analysis demonstrated that stenosis severity, LAP, and PR were predictive of abnormal lesion FFRct, while SC was not. In multivariable analyses considering stenosis severity, presence of PR, LAP, and PR and/or LAP were independently associated with abnormal FFRct: Odds ratio 1.58, 1.68, and 1.53, respectively (p â< â0.02 for all). The presence of PR and/or LAP increased the frequency of abnormal FFRct with mild stenosis (p â< â0.05) with a similar trend with 70-89% stenosis. The combination of 2 HRP (LAP and PR) identified more lesions with FFR < 0.80 than only 1 HRP. CONCLUSIONS: Semiquantitative visual assessment of high-risk plaque features may improve the selection of studies for FFRct.
Assuntos
Doença da Artéria Coronariana , Estenose Coronária , Reserva Fracionada de Fluxo Miocárdico , Angiografia por Tomografia Computadorizada , Constrição Patológica , Angiografia Coronária , Doença da Artéria Coronariana/diagnóstico por imagem , Estenose Coronária/diagnóstico por imagem , Vasos Coronários/diagnóstico por imagem , Humanos , Valor Preditivo dos Testes , Índice de Gravidade de Doença , Tomografia Computadorizada por Raios XRESUMO
BACKGROUND: High amounts of coronary artery calcium (CAC) pose challenges in interpretation of coronary CT angiography (CCTA). The accuracy of stenosis assessment by CCTA in patients with very extensive CAC is uncertain. METHODS: Retrospective study was performed including patients who underwent clinically directed CCTA with CAC score >1000 and invasive coronary angiography within 90 days. Segmental stenosis on CCTA was graded by visual inspection with two-observer consensus using categories of 0%, 1-24%, 25-49%, 50-69%, 70-99%, 100% stenosis, or uninterpretable. Blinded quantitative coronary angiography (QCA) was performed on all segments with stenosis ≥25% by CCTA. The primary outcome was vessel-based agreement between CCTA and QCA, using significant stenosis defined by diameter stenosis ≥70%. Secondary analyses on a per-patient basis and inclusive of uninterpretable segments were performed. RESULTS: 726 segments with stenosis ≥25% in 346 vessels within 119 patients were analyzed. Median coronary calcium score was 1616 (1221-2118). CCTA identification of QCA-based stenosis resulted in a per-vessel sensitivity of 79%, specificity of 75%, positive predictive value (PPV) of 45%, negative predictive value (NPV) of 93%, and accuracy 76% (68 false positive and 15 false negative). Per-patient analysis had sensitivity 94%, specificity 55%, PPV 63%, NPV 92%, and accuracy 72% (30 false-positive and 3 false-negative). Inclusion of uninterpretable segments had variable effect on sensitivity and specificity, depending on whether they are considered as significant or non-significant stenosis. CONCLUSIONS: In patients with very extensive CAC (>1000 Agatston units), CCTA retained a negative predictive value â> â90% to identify lack of significant stenosis on a per-vessel and per-patient level, but frequently overestimated stenosis.
Assuntos
Doença da Artéria Coronariana , Estenose Coronária , Cálcio , Angiografia por Tomografia Computadorizada , Angiografia Coronária , Doença da Artéria Coronariana/diagnóstico por imagem , Estenose Coronária/diagnóstico por imagem , Humanos , Valor Preditivo dos Testes , Estudos RetrospectivosRESUMO
BACKGROUND AND AIMS: We sought to assess the performance of a comprehensive machine learning (ML) risk score integrating circulating biomarkers and computed tomography (CT) measures for the long-term prediction of hard cardiac events in asymptomatic subjects. METHODS: We studied 1069 subjects (age 58.2 ± 8.2 years, 54.0% males) from the prospective EISNER trial who underwent coronary artery calcium (CAC) scoring CT, serum biomarker assessment, and long-term follow-up. Epicardial adipose tissue (EAT) was quantified from CT using fully automated deep learning software. Forty-eight serum biomarkers, both established and novel, were assayed. An ML algorithm (XGBoost) was trained using clinical risk factors, CT measures (CAC score, number of coronary lesions, aortic valve calcium score, EAT volume and attenuation), and circulating biomarkers, and validated using repeated 10-fold cross validation. RESULTS: At 14.5 ± 2.0 years, there were 50 hard cardiac events (myocardial infarction or cardiac death). The ML risk score (area under the receiver operator characteristic curve [AUC] 0.81) outperformed the CAC score (0.75) and ASCVD risk score (0.74; both p = 0.02) for the prediction of hard cardiac events. Serum biomarkers provided incremental prognostic value beyond clinical data and CT measures in the ML model (net reclassification index 0.53 [95% CI: 0.23-0.81], p < 0.0001). Among novel biomarkers, MMP-9, pentraxin 3, PIGR, and GDF-15 had highest variable importance for ML and reflect the pathways of inflammation, extracellular matrix remodeling, and fibrosis. CONCLUSIONS: In this prospective study, ML integration of novel circulating biomarkers and noninvasive imaging measures provided superior long-term risk prediction for cardiac events compared to current risk assessment tools.
Assuntos
Doença da Artéria Coronariana , Calcificação Vascular , Idoso , Biomarcadores , Angiografia Coronária , Doença da Artéria Coronariana/diagnóstico por imagem , Vasos Coronários/diagnóstico por imagem , Feminino , Humanos , Aprendizado de Máquina , Masculino , Pessoa de Meia-Idade , Valor Preditivo dos Testes , Estudos Prospectivos , Medição de Risco , Fatores de RiscoRESUMO
Myocardial perfusion imaging with single photon emission computed tomography or positron emission tomography is commonly used for diagnosis and risk stratification in patients with known or suspected coronary artery disease. Current scanners often incorporate computed tomography for attenuation correction, resulting in a wealth of clinical and imaging information associated with a typical study. Novel highly efficient artificial intelligence (AI) tools have emerged, revolutionizing image analysis with direct and accurate extraction of information from cardiovascular images. These methods have accuracy similar or better to expert interpretation, without the need for timely manual adjustments or measurements. Additionally, artificial intelligence-based algorithms have been developed to integrate the large volume of clinical and imaging information to improve disease diagnosis and risk estimation. Lastly, explainable AI techniques are being developed, overcoming the traditional perception of AI as a "black box" by presenting the rationale for the computed decision or recommendation through attention maps and individualized explanations of risk estimates. In this review we focus on these applications of the latest AI tools in nuclear cardiology and non-contrast cardiac CT.