Deep Learning for Automatic Calcium Scoring in Population-Based Cardiovascular Screening

The Role of Artificial Intelligence in Coronary Calcium Scoring in Standard Cardiac Computed Tomography and Chest Computed Tomography With Different Reconstruction Kernels

PURPOSE

To assess the correlation of coronary calcium score (CS) obtained by artificial intelligence (AI) with those obtained by electrocardiography gated standard cardiac computed tomography (CCT) and nongated chest computed tomography (ChCT) with different reconstruction kernels.

PATIENTS AND METHODS

Seventy-six patients received standard CCT and ChCT simultaneously. We compared CS obtained in 4 groups: CS CCT , by the traditional method from standard CCT, 25 cm field of view, 3 mm slice thickness, and kernel filter convolution 12 (FC12); CS AICCT , by AI from the standard CCT; CS ChCTsoft , by AI from the non-gated CCT, 40 cm field of view, 3 mm slice thickness, and a soft kernel FC02; and CS ChCTsharp , by AI from CCT image with same parameters for CS ChCTsoft except for using a sharp kernel FC56. Statistical analyses included Spearman rank correlation coefficient (ρ), intraclass correlation (ICC), Bland-Altman plots, and weighted kappa analysis (κ).

RESULTS

The CS AICCT was consistent with CS CCT (ρ = 0.994 and ICC of 1.00, P < 0.001) with excellent agreement with respect to cardiovascular (CV) risk categories of the Agatston score (κ = 1.000). The correlation between CS ChCTsoft and CS ChCTsharp was good (ρ = 0.912, 0.963 and ICC = 0.929, 0.948, respectively, P < 0.001) with a tendency of underestimation (Bland-Altman mean difference and 95% upper and lower limits of agreements were 329.1 [-798.9 to 1457] and 335.3 [-651.9 to 1322], respectively). The CV risk category agreement between CS ChCTsoft and CS ChCTsharp was moderate (κ = 0.556 and 0.537, respectively).

CONCLUSIONS

There was an excellent correlation between CS CCT and CS AICCT , with excellent agreement between CV risk categories. There was also a good correlation between CS CCT and CS obtained by ChCT albeit with a tendency for underestimation and moderate accuracy in terms of CV risk assessment.

Long-term outcomes of lung cancer screening in males and females

BACKGROUND

This study explored female and male overall mortality and lung cancer (LC) survival in two LC screening (LCS) populations, focusing on the predictive value of coronary artery calcification (CAC) at baseline low-dose computed tomography (LDCT).

METHODS

This retrospective study analysed data of 6495 heavy smokers enrolled in the MILD and BioMILD LCS trials between 2005 and 2016. The primary objective of the study was to assess sex differences in all-cause mortality and LC survival. CAC scores were automatically calculated on LDCT images by a validated artificial intelligence (AI) software. Sex differences in 12-year cause-specific mortality rates were stratified by age, pack-years and CAC score.

RESULTS

The study included 2368 females and 4127 males. The 12-year all-cause mortality rates were 4.1 % in females and 7.7 % in males (p < 0.0001), and median CAC score was 8.7 vs. 41 respectively (p < 0.0001). All-cause mortality increased with rising CAC scores (log-rank test, p < 0.0001) for both sexes. Although LC incidence was not different between the two sexes, females had lower rates of 12-year LC mortality (1.0 % vs. 1.9 %, p = 0.0052), and better LC survival from diagnosis (72.3 % vs. 51.7 %; p = 0.0005), with a similar proportion of stage I (58.1 % vs. 51.2 %, p = 0.2782).

CONCLUSIONS

Our findings demonstrate that female LCS participants had lower rates of all-cause mortality at 12 years and better LC survival than their male counterparts, with similar LC incidence rates and stage at diagnosis. The lower CAC burden observed in women at all ages might contribute to explain their lower rates of all-cause mortality and better LC survival.

Artificial intelligence using a deep learning versus expert computed tomography human reading in calcium score and coronary artery calcium data and reporting system classification

BACKGROUND

Artificial intelligence (AI) applied to cardiac imaging may provide improved processing, reading precision and advantages of automation. Coronary artery calcium (CAC) score testing is a standard stratification tool that is rapid and highly reproducible. We analyzed CAC results of 100 studies in order to determine the accuracy and correlation between the AI software (Coreline AVIEW, Seoul, South Korea) and expert level-3 computed tomography (CT) human CAC interpretation and its performance when coronary artery disease data and reporting system (coronary artery calcium data and reporting system) classification is applied.

METHODS

A total of 100 non-contrast calcium score images were selected by blinded randomization and processed with the AI software versus human level-3 CT reading. The results were compared and the Pearson correlation index was calculated. The CAC-DRS classification system was applied, and the cause of category reclassification was determined using an anatomical qualitative description by the readers.

RESULTS

The mean age was age 64.5 years, with 48% female. The absolute CAC scores between AI versus human reading demonstrated a highly significant correlation (Pearson coefficient R  = 0.996); however, despite these minimal CAC score differences, 14% of the patients had their CAC-DRS category reclassified. The main source of reclassification was observed in CAC-DRS 0-1, where 13 were recategorized, particularly between studies having a CAC Agatston score of 0 versus 1. Qualitative description of the errors showed that the main cause of misclassification was AI underestimation of right coronary calcium, AI overestimation of right ventricle densities and human underestimation of right coronary artery calcium.

CONCLUSION

Correlation between AI and human values is excellent with absolute numbers. When the CAC-DRS classification system was adopted, there was a strong correlation in the respective categories. Misclassified were predominantly in the category of CAC = 0, most often with minimal values of calcium volume. Additional algorithm optimization with enhanced sensitivity and specificity for low values of calcium volume will be required to enhance AI CAC score utilization for minimal disease. Over a broad range of calcium scores, AI software for calcium scoring had an excellent correlation compared to human expert reading and in rare cases determined calcium missed by human interpretation.

Automated Coronary Artery Calcium and Quantitative Emphysema in Lung Cancer Screening: Association With Mortality, Lung Cancer Incidence, and Airflow Obstruction

PURPOSE

To assess automated coronary artery calcium (CAC) and quantitative emphysema (percentage of low attenuation areas [%LAA]) for predicting mortality and lung cancer (LC) incidence in LC screening. To explore correlations between %LAA, CAC, and forced expiratory value in 1 second (FEV 1 ) and the discriminative ability of %LAA for airflow obstruction.

MATERIALS AND METHODS

Baseline low-dose computed tomography scans of the BioMILD trial were analyzed using an artificial intelligence software. Univariate and multivariate analyses were performed to estimate the predictive value of %LAA and CAC. Harrell C -statistic and time-dependent area under the curve (AUC) were reported for 3 nested models (Model survey : age, sex, pack-years; Model survey-LDCT : Model survey plus %LAA plus CAC; Model final : Model survey-LDCT plus selected confounders). The correlations between %LAA, CAC, and FEV 1 and the discriminative ability of %LAA for airflow obstruction were tested using the Pearson correlation coefficient and AUC-receiver operating characteristic curve, respectively.

RESULTS

A total of 4098 volunteers were enrolled. %LAA and CAC independently predicted 6-year all-cause (Model final hazard ratio [HR], 1.14 per %LAA interquartile range [IQR] increase [95% CI, 1.05-1.23], 2.13 for CAC ≥400 [95% CI, 1.36-3.28]), noncancer (Model final HR, 1.25 per %LAA IQR increase [95% CI, 1.11-1.37], 3.22 for CAC ≥400 [95%CI, 1.62-6.39]), and cardiovascular (Model final HR, 1.25 per %LAA IQR increase [95% CI, 1.00-1.46], 4.66 for CAC ≥400, [95% CI, 1.80-12.58]) mortality, with an increase in concordance probability in Model survey-LDCT compared with Model survey ( P <0.05). No significant association with LC incidence was found after adjustments. Both biomarkers negatively correlated with FEV 1 ( P <0.01). %LAA identified airflow obstruction with a moderate discriminative ability (AUC, 0.738).

CONCLUSIONS

Automated CAC and %LAA added prognostic information to age, sex, and pack-years for predicting mortality but not LC incidence in an LC screening setting. Both biomarkers negatively correlated with FEV 1 , with %LAA enabling the identification of airflow obstruction with moderate discriminative ability.

Fully automated deep learning powered calcium scoring in patients undergoing myocardial perfusion imaging

BACKGROUND

To assess the accuracy of fully automated deep learning (DL) based coronary artery calcium scoring (CACS) from non-contrast computed tomography (CT) as acquired for attenuation correction (AC) of cardiac single-photon-emission computed tomography myocardial perfusion imaging (SPECT-MPI).

METHODS AND RESULTS

Patients were enrolled in this study as part of a larger prospective study (NCT03637231). In this study, 56 Patients who underwent cardiac SPECT-MPI due to suspected coronary artery disease (CAD) were prospectively enrolled. All patients underwent non-contrast CT for AC of SPECT-MPI twice. CACS was manually assessed (serving as standard of reference) on both CT datasets (n = 112) and by a cloud-based DL tool. The agreement in CAC scores and CAC score risk categories was quantified. For the 112 scans included in the analysis, interscore agreement between the CAC scores of the standard of reference and the DL tool was 0.986. The agreement in risk categories was 0.977 with a reclassification rate of 3.6%. Heart rate, image noise, body mass index (BMI), and scan did not significantly impact (p=0.09 - p=0.76) absolute percentage difference in CAC scores.

CONCLUSION

A DL tool enables a fully automated and accurate estimation of CAC scores in patients undergoing non-contrast CT for AC of SPECT-MPI.

Fully automated calcium scoring predicts all-cause mortality at 12 years in the MILD lung cancer screening trial

Coronary artery calcium (CAC) is a known risk factor for cardiovascular (CV) events and mortality but is not yet routinely evaluated in low-dose computed tomography (LDCT)-based lung cancer screening (LCS). The present analysis explored the capacity of a fully automated CAC scoring to predict 12-year mortality in the Multicentric Italian Lung Detection (MILD) LCS trial. The study included 2239 volunteers of the MILD trial who underwent a baseline LDCT from September 2005 to January 2011, with a median follow-up of 190 months. The CAC score was measured by a commercially available fully automated artificial intelligence (AI) software and stratified into five strata: 0, 1-10, 11-100, 101-400, and > 400. Twelve-year all-cause mortality was 8.5% (191/2239) overall, 3.2% with CAC = 0, 4.9% with CAC = 1-10, 8.0% with CAC = 11-100, 11.5% with CAC = 101-400, and 17% with CAC > 400. In Cox proportional hazards regression analysis, CAC > 400 was associated with a higher 12-year all-cause mortality both in a univariate model (hazard ratio, HR, 5.75 [95% confidence interval, CI, 2.08-15.92] compared to CAC = 0) and after adjustment for baseline confounders (HR, 3.80 [95%CI, 1.35-10.74] compared to CAC = 0). All-cause mortality significantly increased with increasing CAC (7% in CAC ≤ 400 vs. 17% in CAC > 400, Log-Rank p-value <0.001). Non-cancer at 12 years mortality was 3% (67/2239) overall, 0.8% with CAC = 0, 1.0% with CAC = 1-10, 2.9% with CAC = 11-100, 3.6% with CAC = 101-400, and 8.2% with CAC > 400 (Grey's test p < 0.001). In Fine and Gray's competing risk model, CAC > 400 predicted 12-year non-cancer mortality in a univariate model (sub-distribution hazard ratio, SHR, 10.62 [95% confidence interval, CI, 1.43-78.98] compared to CAC = 0), but the association was no longer significant after adjustment for baseline confounders. In conclusion, fully automated CAC scoring was effective in predicting all-cause mortality at 12 years in a LCS setting.

Evaluation of an artificial intelligence coronary artery calcium scoring model from computed tomography

OBJECTIVES

Coronary artery calcium (CAC) scores derived from computed tomography (CT) scans are used for cardiovascular risk stratification. Artificial intelligence (AI) can assist in CAC quantification and potentially reduce the time required for human analysis. This study aimed to develop and evaluate a fully automated model that identifies and quantifies CAC.

METHODS

Fully convolutional neural networks for automated CAC scoring were developed and trained on 2439 cardiac CT scans and validated using 771 scans. The model was tested on an independent set of 1849 cardiac CT scans. Agatston CAC scores were further categorised into five risk categories (0, 1-10, 11-100, 101-400, and > 400). Automated scores were compared to the manual reference standard (level 3 expert readers).

RESULTS

Of 1849 scans used for model testing (mean age 55.7 ± 10.5 years, 49% males), the automated model detected the presence of CAC in 867 (47%) scans compared with 815 (44%) by human readers (p = 0.09). CAC scores from the model correlated very strongly with the manual score (Spearman's r = 0.90, 95% confidence interval [CI] 0.89-0.91, p < 0.001 and intraclass correlation coefficient = 0.98, 95% CI 0.98-0.99, p < 0.001). The model classified 1646 (89%) into the same risk category as human observers. The Bland-Altman analysis demonstrated little difference (1.69, 95% limits of agreement: -41.22, 44.60) and there was almost excellent agreement (Cohen's κ = 0.90, 95% CI 0.88-0.91, p < 0.001). Model analysis time was 13.1 ± 3.2 s/scan.

CONCLUSIONS

This artificial intelligence-based fully automated CAC scoring model shows high accuracy and low analysis times. Its potential to optimise clinical workflow efficiency and patient outcomes requires evaluation.

KEY POINTS

• Coronary artery calcium (CAC) scores are traditionally assessed using cardiac computed tomography and require manual input by human operators to identify calcified lesions. • A novel artificial intelligence (AI)-based model for fully automated CAC scoring was developed and tested on an independent dataset of computed tomography scans, showing very high levels of correlation and agreement with manual measurements as a reference standard. • AI has the potential to assist in the identification and quantification of CAC, thereby reducing the time required for human analysis.

Opportunistic deep learning powered calcium scoring in oncologic patients with very high coronary artery calcium (≥ 1000) undergoing 18F-FDG PET/CT

Our aim was to identify and quantify high coronary artery calcium (CAC) with deep learning (DL)-powered CAC scoring (CACS) in oncological patients with known very high CAC (≥ 1000) undergoing 18F-FDG-PET/CT for re-/staging. 100 patients were enrolled: 50 patients with Agatston scores ≥ 1000 (high CACS group), 50 patients with Agatston scores < 1000 (negative control group). All patients underwent oncological 18F-FDG-PET/CT and cardiac SPECT myocardial perfusion imaging (MPI) by 99mTc-tetrofosmin within 6 months. CACS was manually performed on dedicated non-contrast ECG-gated CT scans obtained from SPECT-MPI (reference standard). Additionally, CACS was performed fully automatically with a user-independent DL-CACS tool on non-contrast, free-breathing, non-gated CT scans from 18F-FDG-PET/CT examinations. Image quality and noise of CT scans was assessed. Agatston scores obtained by manual CACS and DL tool were compared. The high CACS group had Agatston scores of 2200 ± 1620 (reference standard) and 1300 ± 1011 (DL tool, average underestimation of 38.6 ± 26%) with an intraclass correlation of 0.714 (95% CI 0.546, 0.827). Sufficient image quality significantly improved the DL tool's capability of correctly assigning Agatston scores ≥ 1000 (p = 0.01). In the control group, the DL tool correctly assigned Agatston scores < 1000 in all cases. In conclusion, DL-based CACS performed on non-contrast free-breathing, non-gated CT scans from 18F-FDG-PET/CT examinations of patients with known very high (≥ 1000) CAC underestimates CAC load, but correctly assigns an Agatston scores ≥ 1000 in over 70% of cases, provided sufficient CT image quality. Subgroup analyses of the control group showed that the DL tool does not generate false-positives.

Association of Sinoatrial Node Radiation Dose With Atrial Fibrillation and Mortality in Patients With Lung Cancer

Importance

Atrial fibrillation (AF) can develop following thoracic irradiation. However, the critical cardiac substructure responsible for AF has not been properly studied.

Objective

To describe the incidence of AF in patients with lung cancer and determine predictive cardiac dosimetric parameters.

Design, Setting, and Participants

This retrospective cohort study was performed at a single referral center and included 239 patients diagnosed with limited-stage small cell lung cancer (SCLC) and 321 patients diagnosed with locally advanced non-small cell lung cancer (NSCLC) between August 2008 and December 2019 who were treated with definitive chemoradiotherapy.

Exposures

Radiation dose exposure to cardiac substructures, including the chambers, coronary arteries, and cardiac conduction nodes, were calculated for each patient.

Main Outcomes and Measures

Main outcomes were AF and overall survival.

Results

Of the 239 and 321 patients with SCLC and NSCLC, the median (IQR) age was 68 (60-73) years and 67 (61-75) years, and 207 (86.6%) and 261 (81.3%) were men, respectively. At a median (IQR) follow-up time of 32.7 (22.1-56.6) months, 9 and 17 patients experienced new-onset AF in the SCLC and NSCLC cohorts, respectively. The maximum dose delivered to the sinoatrial node (SAN Dmax) exhibited the highest predictive value for prediction of AF. A higher SAN Dmax significantly predicted an increased risk of AF in patients with SCLC (adjusted hazard ratio [aHR], 14.91; 95% CI, 4.00-55.56; P < .001) and NSCLC (aHR, 15.67; 95% CI, 2.08-118.20; P = .008). However, SAN Dmax was not associated with non-AF cardiac events. Increased SAN Dmax was significantly associated with poor overall survival in patients with SCLC (aHR, 2.68; 95% CI, 1.53-4.71; P < .001) and NSCLC (aHR, 1.97; 95% CI, 1.45-2.68; P < .001).

Conclusions and Relevance

In this cohort study, results suggest that incidental irradiation of the SAN during chemoradiotherapy may be associated with the development of AF and increased mortality. This supports the need to minimize radiation dose exposure to the SAN during radiotherapy planning and to consider close follow-up for the early detection of AF in patients receiving thoracic irradiation.

Risk stratification using coronary artery calcium scoring based on low tube voltage computed tomography

To determine if coronary artery calcium (CAC) scoring using computed tomography at 80 kilovolt-peak (kVp) and 70-kVp and tube voltage-adapted scoring-thresholds allow for accurate risk stratification as compared to the standard 120-kVp protocol. We prospectively included 170 patients who underwent standard CAC scanning at 120-kVp and 200 milliamperes and additional scans with 80-kVp and 70-kVp tube voltage with adapted tube current to normalize image noise across scans. Novel kVp-adapted thresholds were applied to calculate CAC scores from the low-kVp scans and were compared to those from standard 120-kVp scans by assessing risk reclassification rates and agreement using Kendall's rank correlation coefficients (Τb) for risk categories bounded by 0, 1, 100, and 400. Interreader reclassification rates for the 120-kVp scans were assessed. Agreement for risk classification obtained from 80-kVp and 70-kVp scans as compared to 120-kVp was good (Τb = 0.967 and 0.915, respectively; both p < 0.001) with reclassification rates of 7.1% and 17.2%, respectively, mostly towards a lower risk category. By comparison, the interreader reclassification rate was 4.1% (Τb = 0.980, p < 0.001). Reclassification rates were dependent on body mass index (BMI) with 7.1% and 13.6% reclassifications for the 80-kVp and 70-kVp scans, respectively, in patients with a BMI < 30 kg/m2 (n = 140), and 2.9% and 7.4%, respectively, in patients with a BMI < 25 kg/m2 (n = 68). Mean effective radiation dose from the 120-kVp, the 80-kVp, and 70-kVp scans was 0.54 ± 0.03, 0.42 ± 0.02, and 0.26 ± 0.02 millisieverts. CAC scoring with reduced tube voltage allows for accurate risk stratification if kVp-adapted thresholds for calculation of CAC scores are applied.ClinicalTrials.gov NCT03637231.

Validation of deep learning-based fully automated coronary artery calcium scoring using non-ECG-gated chest CT in patients with cancer

This study aimed to demonstrate clinical feasibility of deep learning (DL)-based fully automated coronary artery calcium (CAC) scoring software using non-electrocardiogram (ECG)-gated chest computed tomography (CT) from patients with cancer. Overall, 913 patients with colorectal or gastric cancer who underwent non-contrast-enhanced chest CT between 2013 and 2015 were included. Agatston scores obtained by manual segmentation of CAC on chest CT were used as reference. Reliability of automated CAC score acquisition was evaluated using intraclass correlation coefficients (ICCs). The agreement for cardiovascular disease (CVD) risk stratification was assessed with linearly weighted k statistics. ICCs between the manual and automated CAC scores were 0.992 (95% CI, 0.991 and 0.993, p<0.001) for total Agatston scores, 0.863 (95% CI, 0.844 and 0.880, p<0.001) for the left main, 0.964 (95% CI, 0.959 and 0.968, p<0.001) for the left anterior descending, 0.962 (95% CI, 0.956 and 0.966, p<0.001) for the left circumflex, and 0.980 (95% CI, 0.978 and 0.983, p<0.001) for the right coronary arteries. The agreement for cardiovascular risk was excellent (k=0.946, p<0.001). Current DL-based automated CAC software showed excellent reliability for Agatston score and CVD risk stratification using non-ECG gated CT scans and might allow the identification of high-risk cancer patients for CVD.

Diagnostic Value of Fully Automated Artificial Intelligence Powered Coronary Artery Calcium Scoring from 18F-FDG PET/CT

Objectives: The objective of this study was to assess the feasibility and accuracy of a fully automated artificial intelligence (AI) powered coronary artery calcium scoring (CACS) method on ungated CT in oncologic patients undergoing 18F-FDG PET/CT. Methods: A total of 100 oncologic patients examined between 2007 and 2015 were retrospectively included. All patients underwent 18F-FDG PET/CT and cardiac SPECT myocardial perfusion imaging (MPI) by 99mTc-tetrofosmin within 6 months. CACS was manually performed on non-contrast ECG-gated CT scans obtained from SPECT-MPI (i.e., reference standard). Additionally, CACS was performed using a cloud-based, user-independent tool (AI-CACS) on ungated CT scans from 18F-FDG-PET/CT examinations. Agatston scores from the manual CACS and AI-CACS were compared. Results: On a per-patient basis, the AI-CACS tool achieved a sensitivity and specificity of 85% and 90% for the detection of CAC. Interscore agreement of CACS between manual CACS and AI-CACS was 0.88 (95% CI: 0.827, 0.918). Interclass agreement of risk categories was 0.8 in weighted Kappa analysis, with a reclassification rate of 44% and an underestimation of one risk category by AI-CACS in 39% of cases. On a per-vessel basis, interscore agreement of CAC scores ranged from 0.716 for the circumflex artery to 0.863 for the left anterior descending artery. Conclusions: Fully automated AI-CACS as performed on non-contrast free-breathing, ungated CT scans from 18F-FDG-PET/CT examinations is feasible and provides an acceptable to good estimation of CAC burden. CAC load on ungated CT is, however, generally underestimated by AI-CACS, which should be taken into account when interpreting imaging findings.

Outstanding negative prediction performance of solid pulmonary nodule volume AI for ultra-LDCT baseline lung cancer screening risk stratification

OBJECTIVE

To evaluate performance of AI as a standalone reader in ultra-low-dose CT lung cancer baseline screening, and compare it to that of experienced radiologists.

METHODS

283 participants who underwent a baseline ultra-LDCT scan in Moscow Lung Cancer Screening, between February 2017-2018, and had at least one solid lung nodule, were included. Volumetric nodule measurements were performed by five experienced blinded radiologists, and independently assessed using an AI lung cancer screening prototype (AVIEW LCS, v1.0.34, Coreline Soft, Co. ltd, Seoul, Korea) to automatically detect, measure, and classify solid nodules. Discrepancies were stratified into two groups: positive-misclassification (PM); nodule classified by the reader as a NELSON-plus /EUPS-indeterminate/positive nodule, which at the reference consensus read was < 100 mm3, and negative-misclassification (NM); nodule classified as a NELSON-plus /EUPS-negative nodule, which at consensus read was ≥ 100 mm3.

RESULTS

1149 nodules with a solid-component were detected, of which 878 were classified as solid nodules. For the largest solid nodule per participant (n = 283); 61 [21.6 %; 53 PM, 8 NM] discrepancies were reported for AI as a standalone reader, compared to 43 [15.1 %; 22 PM, 21 NM], 36 [12.7 %; 25 PM, 11 NM], 29 [10.2 %; 25 PM, 4 NM], 28 [9.9 %; 6 PM, 22 NM], and 50 [17.7 %; 15 PM, 35 NM] discrepancies for readers 1, 2, 3, 4, and 5 respectively.

CONCLUSION

Our results suggest that through the use of AI as an impartial reader in baseline lung cancer screening, negative-misclassification results could exceed that of four out of five experienced radiologists, and radiologists' workload could be drastically diminished by up to 86.7%.