Current guidelines recommend estimating 10-year risk of major adverse cardiovascular events (MACE) to establish statin candidacy for the primary prevention of atherosclerotic cardiovascular disease (ASCVD). The current ASCVD risk score requires age, sex, race, systolic blood pressure, hypertension treatment, smoking, type 2 diabetes and a lipid panel. As these variables are often not available in the electronic record, other approaches for population-based screening are desirable. Here, we developed a deep learning model (CXR-CVD risk), to estimate 10-year cardiovascular risk from a routine chest radiograph (CXR).

The CXR-CVD risk model was developed using 147,497 CXRs of 40,643 participants from the PLCO cancer screening trial. The model was trained to predict cardiovascular mortality from a single CXR image. Independent testing was performed in a second separate cohort of 11,430 outpatients potentially eligible for primary prevention (low-density lipoprotein cholesterol 70-190 mg/dl, no prevalent diabetes and no prior MACE). Statin eligibility was defined as a 10-year MACE risk =7.5%. The prognostic value of CXR-CVD risk was compared to the established ASCVD risk score in the subset of 2,401 (21%) where the variables necessary to calculate ASCVD risk were available. The primary outcome was observed 10-year incident MACE (stroke and myocardial infarction). Hazard ratios and c-statistics for MACE were estimated using Cox proportional hazards regression.

In the independent testing dataset of 11,430 patients (mean age 60.1±6.7 years; 42.9% male), 1096 (9.6%) MACE occurred over median follow-up of 10.3 years. There was a significant association of CXR-CVD risk and MACE in statin eligible patients (HR: 2.03 [1.81-2.30]; p<0.001), which remained significant after adjustment for cardiovascular risk factors (adjusted HR: 1.63 [1.43-1.86]; p<0.001). In the subgroup where all variables necessary to calculate ASCVD risk were available, the performance of CXR-CVD risk was similar (c-statistic 0.64 vs. 0.65; p=0.48) to and additive to the ASCVD risk score (adjusted HR: 1.58 [1.20-2.09]; p=0.001).

Based on a single routine CXR image, our deep learning model predict10-year incident MACE with similar performance and incremental to the established clinical standard. ACXR imageare commonly available, our approach may help identify individualat high risk for cardiovascular disease, prompting risk factor assessment and targeted prevention.

Deep learning can estimate cardiovascular risk from a routine CXR image similar to the current clinical standard. This enables opportunistic screening to identify high-risk patients who would benefit from prevention with statins.

To investigate the clinical impact of implementing an artificial intelligence-based computer-aided detection system (AI-CAD) for chest X-rays (CXRs) via comparison of referral rate to chest CTs before and after the implementation.

A commercialized AI-CAD that can identify pulmonary nodule, infiltration, and pneumothorax in a CXR was implemented for daily clinical practice in a tertiary referral institution in January 2020. After the implementation, referring clinicians decided to request whether CXR interpretation with or without AI-CAD. All CXRs were formally interpreted by radiologists, while AI-CAD was used only when examinations with AI-CAD were requested. For the analysis, we included all CXR examinations obtained from patients who visited pulmonology outpatient clinics from January to December 2019 (before AI-CAD implementation) and from January to December 2020 (after AI-CAD implementation). Both CXRs interpreted with and without AI-CAD were included for CXRs after AI-CAD implementation. We compared the following metrics between CXRs obtained before and after AI-CAD implementation: 1) CT referral rate (Proportion of CXRs after which chest CT was obtained between 1 to 30 days), 2) true-positive referral rate (Proportion of CXRs with identification of abnormal finding followed by the presence of abnormality on chest CT), and 3) False-positive referral rate (Proportion of CXRs with identification of abnormal finding followed by the absence of abnormality on chest CT).

28,546 CXRs (from 14,565 patients; 49% men; mean age 66 years) were included before the AI-CAD implementation, while 25,888 CXRs (from 12,928 patients; 50% men; mean age 66 years; AI-CAD used in 13%) were included after the AI-CAD implementation. The CT referral rate was significantly increased after the AI-CAD implementation (7.4% to 8.3%; P<.001). The false-positive referral rate was significantly increased after the AI-CAD implementation (0.50% to 0.64%; P=.039), while the true-positive referral rate did not change significantly (0.98% to 1.07%; P=.282).

Even though the AI-CAD wapartly applied, requestfor chest CT and false-positive referralwere significantly increased after the implementation of AI-CAD for CXR interpretation in daily clinical practice.

Increased referral to chest CT examination and false-positive identification of abnormalities could be important drawbacks of clinical implementation of an AI-CAD for CXRs

Chest X-ray (CXR) is a commonly used imaging modality that accounts for over 20% of the diagnostic errors in radiology, with radiologists spending the majority of their time to generate reports. To address this challenge, we developed a new deep learning-based model to automatically detect multiple abnormalities and generate diagnostic reports.

A Multi-task, Optimal-recommendation, and Max-predictive Classification and Segmentation (MOM ClaSeg) model integrating Mask R-CNN deep learning and Decision Fusion Networks is developed using 310,333 confirmed adult CXR images collected from multiple hospitals around the world. The image set contains 243,262 abnormal images depicting 65 different abnormalities and 67,071 normal images. Unlike traditional CAD models that only detect a single type of abnormality, this new model is optimally trained to detect multiple abnormalities of different types of diseases visible in CXR images and generates diagnostic reports of radiological impression for all detected abnormalities using a bifurcated gradient boosted decision tree. Then, an independent testing dataset including 22,642 CXR images collected from 54 hospitals is processed by the new model. The Multi Disease Detection with Optimal Recommendation (MDD-OR) reports of these images are generated and compared with single disease detection (SDD) results. A panel of 8 radiologists read these images and evaluate the new model performance.

Radiologists detected 43 different types of abnormalities distributed on 6,656 regions of interest (ROIs) from 4,068 CXR images. Comparing with radiologists’ detection results, our MDD-OR reports detected 6,009 true-positive ROIs and 6,379 false positive ROIs, while traditional SDD reports detected 5,012 TP ROIs and 14,492 FP ROIs. Thus, using the new model increases sensitivity from 75.3% to 90.3% and reduces FP rate from 0.64 to 0.28 per image. In addition, 70.44% of MDD-OR first choice reports are directly accepted by radiologists, 19.81% of reports are selected by radiologists from the top 2-6 MDD-OR recommendations, and 9.75% of reports from the top 7 and beyond MDD-OR recommendation.

Thistudy presentthe first AI-based model that automatically detect65 different abnormalitiein CXR imageand generatediagnostic reports. Our independent observer evaluation study ia promising step toward improved diagnostic performance and productivity of radiologistin future clinical practice.

This study develops and tests the first AI-based model that can detect multiple abnormalities and generate diagnostic reports to help improve the diagnostic accuracy and efficiency of radiologists.Please visit the Learning Center to also view this presentation in hardcopy format.

To observe how artificial intelligence (AI) affects the reading times of radiologists for the daily interpretation of chest radiographs.

Radiologists who agreed to be collected the reading times of their daily chest radiograph interpretations from September to December 2021 were prospectively recruited. Reading time was defined as the duration in seconds from opening the radiograph to transcribing the opened image by the same radiologist on a picture archiving and communicating system (PACS). As commercially available AI software is integrated to PACS for all chest radiographs in our hospital, radiologists could refer to AI results when interpreting radiographs for two months of the study period (AI-aided). During the other two months, AI results on PACS were automatically blinded to the radiologists (AI-unaided). Outliers in time were excluded using the interquartile range method. The linear mixed model was used to compare reading times considering the random effects of radiologists and patients.

A total of 11 radiologists participated prospectively and 18,884 chest radiographs were included after excluding outliers (>51 seconds). Total reading times significantly shortened with the use of AI compared with AI-unaided interpretations (estimated mean 12.8 sec vs. 14.4 sec, p<0.001). Reading times for outpatients were significantly decreased more than that for inpatients in AI-aided interpretation (difference -2.3 sec vs. -0.5 sec, p<0.001). When there was no abnormality detected by AI on radiographs, reading times were significantly shorter in AI-aided interpretation (estimated mean 10.8 sec vs. 13.1 sec, p<0.001). However, when there was any abnormality detected by AI, reading times were not significantly different between AI-aided and AI-unaided interpretations (estimated mean 18.6 sec vs. 18.4 sec, p=0.452). When the total abnormality score analyzed by AI was considered as a continuous variable, reading times were significantly increased as higher scores, and more significantly increased with the use of AI compared with AI-unaided interpretations (coefficient 0.09 sec vs. 0.06 sec, p<0.001).

The reading time of radiologistfor chest radiographwainfluenced by the availability of AI-based results. Overall reading time shortened when radiologistreferred to AI, but abnormalitiedetected by AI on chest radiographcould lengthen reading times.

This study demonstrated how AI improved efficiency in workflow of radiologists, by showing impact on reading time.

Prior studies have reported a variable but high incidence of suboptimal or poor-quality chest radiographs (CXR). We performed a multicenter (11 hospitals), multinational (7 countries), cross-sectional study to assess the prevalence and causes of poor-quality portable CR.

Our IRB-approved, retrospective study included 8038 CXR from 11 hospitals from 7 countries (Brazil, India, Lebanon, Iran, Italy, Croatia, US). The CXR were uploaded on a secure server-based algorithm (CARPL AI platform) for annotation into the following suboptimal categories. 1. Missing anatomy - left and right costophrenic angles, apices, part of left or right lung. 2. Minor and major rotation. 3. Overlying structures such as chin (in neck flexion), ECG coils, metal or non-metal structures. 4. Artifacts- processing or cassette. 5. Under or over-exposure 6. Low lung volumes. Descriptive statistics was used to analyze the data.

The distribution of suboptimal radiograph varied significantly between each site (p <0.001). There were significant variations in prevalence of optimal CR, with an average optimality rate of 25% (range 8-62% across different sites). Most CR had two or more causes for suboptimality (74%). The most frequent single causes of suboptimality were missed anatomy 9% (n=738 CR, range 4-37%); overlying structures 5% (n=429, range 5-39%); major rotation 7% (n=530, range 5-40%); inadequate inspiration 22% (n=1739, range 20-49%); and under exposure 14% (n=1123, range 0-11%). The distribution of over-exposed CR varied across each site with a range of 1% to 32%.

Our multicenter, multinational study demonstratethat almost three-quarterof portable CXR have poor quality or are suboptimal (range 48% to 92%).

There is a significant concern for suboptimal portable CXR across the globe. Therefore, there is a need for education and quality improvement measures to reduce the frequency of suboptimal, portable CXR.Please visit the Learning Center to also view this presentation in hardcopy format.

Total lung capacity (TLC) has been estimated using chest radiographs based on time-consuming methods such as planimetric techniques and manual measurements. We aimed to develop and validate a deep learning-based, multidimensional model capable of estimating TLC from chest radiographs and demographic variables with high accuracy and reproducibility using multicenter retrospective cohorts.

Our model was pretrained using 50,000 chest CT scans and fine-tuned on 3,523 pairs of posteroanterior chest radiographs and plethysmographic TLC measurements, and it was validated with three separate, retrospective cohorts: 1) external validation set 1 (EVAL1; n=207), and external validation set 2 (EVAL2; n=216) for technical performance; and 2) an idiopathic pulmonary fibrosis (IPF) cohort (n=217) for clinical utility. Technical performance was evaluated using various agreement measures, and clinical utility was assessed in terms of the prognostic value for overall survival using multivariable Cox regression. The feasibility of substituting predicted TLC for forced vital capacity in the GAP (gender, age, and physiologic variables) index was also analyzed.

The mean difference and within-subject standard deviation between observed and predicted TLC were 0.21 L and 0.73 L, respectively, in EVAL1 and -0.29 L and 0.53 L, respectively, in EVAL2. The repeatability coefficient was 2.03 L (95% confidence interval [CI]: 1.86, 2.25 L) for EVAL1 and 1.46 L (95% CI: 1.33, 1.61 L) for EVAL2. In patients with IPF, greater predicted TLC was associated with lower mortality risk (adjusted hazard ratio for 2-year overall survival, 0.51 [95% CI: 0.32, 0.81; P=.004]). The prognostic discrimination of the GAP index was well maintained after substituting predicted TLC for forced vital capacity (C-index pre-substitution, 0.79 [95% CI: 0.73, 0.86]; post-substitution, 0.77 [95% CI: 0.70, 0.82]; P=.29).

A fully-automatic, deep learning-based model could estimate total lung capacity from chest radiographs, and the model output predicted survival in idiopathic pulmonary fibrosis.

A fully-automatic, deep learning-based model could estimate total lung capacity from frontal chest radiographs and demographic variables accurately and reproducibly, and the model output predicted survival in idiopathic pulmonary fibrosis.