Abstract
Background
To investigate the effectiveness of an artificial intelligence (AI)-based computer-aided detection (CAD) system in identifying incidental lung cancer on cardiac computed tomography (CT) scans and to compare its performance with that of radiologists.
Methods
In this retrospective, multicenter study, 652 cardiac CT scans from 581 patients subsequently diagnosed with lung cancer were analyzed. A commercial AI-CAD system was employed to detect pulmonary lesions on cardiac CT. The detection rate of AI-CAD was compared to that of the radiologist, based on the radiology report, as well as to the detection rate when combining AI-CAD and the radiologist. The characteristics of the lesions detected and missed by the radiologist and AI-CAD were compared.
Results
Radiologists and AI-CAD demonstrated similar detection rates for lung cancer (76.2% vs. 77.4%, P = 0.551). However, combining radiologists and AI-CAD significantly improved the detection rate to 90.4% (P < 0.001) compared to that of the radiologist alone. AI-CAD showed a higher detection rate in identifying small, peripheral, and part-solid lesions (all P < 0.001). Furthermore, AI-CAD outperformed radiologists in detecting limited-stage lung cancer (80.3% vs. 74.7%, P = 0.006). Among lung cancer cases missed by radiologists, 94.2% experienced diagnostic delays of > 100 days, with 78.2% leading to stage progression. AI-CAD identified 58.5% of these diagnostic delays.
Conclusion
AI-CAD demonstrated the potential to improve the detection rate of incidental lung cancer by identifying a subset of lesions that were initially overlooked by radiologists on cardiac CT. It exhibited particular strength in identifying early-stage cancers and small, subsolid lesions.
Keywords: Heart, Computed Tomography, Artificial Intelligence, Lung Cancer
Graphical Abstract
INTRODUCTION
Advances in cardiac computed tomography (CT) technology have transformed it into a highly effective clinical modality, offering precise assessments of atherosclerotic coronary plaque, stenosis, and prognostic insights in cardiovascular disease.1,2 Additionally, cardiac CT has become integral in structural heart disease imaging and catheter-based valve replacement procedures.3,4
With the expanding cardiac CT utilization, increasing attention is given to incidental detection of extracardiac findings. The field of view (FOV) in cross-sectional cardiac imaging encompasses the heart as well as adjacent structures, including lungs, bones, and parts of the abdomen. Among various extracardiac findings detectable on cardiac CT, pulmonary nodule is the most commonly identified. Lung cancer represents a critical finding that should not be overlooked when interpreting cardiac CT. Given the overlap of risk factors between coronary artery disease and lung cancer, patients undergoing CT evaluation for one condition may also be at risk for the other.5,6,7 The reported incidental detection rate of lung cancer on cardiac CT is approximately 0.4%, and employing a wide FOV facilitates the identification of such cancers.8,9,10 However, radiologists often overlook incidental lung cancer on cardiac CT, as their primary focus is centered on cardiac findings.11
Meanwhile, a recent advance in artificial intelligence (AI) technology has led to investigation of the possibility of AI assistance in detecting lung nodules on chest CT scans. This innovation holds the promise of improving diagnostic accuracy and reducing radiologist workload. AI software designed for lung nodule detection on chest CT has recently demonstrated sensitivity ranging from 61.6–98.1%.12,13,14,15,16 Several studies show that radiologists achieve higher sensitivity when assisted by AI compared to those of unaided evaluations,16,17,18,19,20,21 with improvement being particularly significant for less experienced radiologists.16,20 Building on these encouraging circumstances, applying AI for lung nodule detection in cardiac CT has the potential to reduce the number of missing incidental lung cancers. However, no studies to date have specifically investigated the utility of artificial intelligence-based computer-aided detection (AI-CAD) for identifying malignant lung nodules on cardiac CT scans.
The purpose of this study was to evaluate the detection performance of an AI-CAD system trained on chest CT when applied to cardiac CT scans with a lung FOV, particularly for incidentally detected lung cancers, using a multicenter dataset. We also aimed to determine whether the incorporation of AI-CAD into cardiac CT interpretation could improve the detection performance for incidental lung cancer compared with radiologist interpretation alone.
METHODS
Patients
The databases of the three institutions were queried to identify patients who underwent cardiac CT scans between 2010 and 2022, before being diagnosed with lung cancer, either histologically or clinically. In total, 1,041 cardiac CT scans from 863 patients were identified.
We excluded 320 scans with no detectable lesions, 34 scans from patients with known extrathoracic malignancy, 13 scans lacking a full FOV covering the entire thorax, eight scans where lesions were indistinguishable owing to lung collapse or effusion, eight scans with lung cancers presented as a mediastinal mass, lymph node enlargement, or bronchial wall thickening without definite lung nodules, three scans with lung cancer presented as multiple (more than eight in number) lung metastases without definite dominant lesions, and three scans that could not be accessed owing to an error in loading Digital Imaging and Communications in Medicine files.
All serial cardiac CT scans from individual patients were included. For cases with multiple lesions diagnosed as primary lung cancers on a single CT scan, all identified lung cancers were included in the analysis. Consequently, up to three lesions per patient were included in the analysis. Ultimately, 652 CT scans from 581 patients, identifying 665 lesions, were included in the study (Fig. 1).
Fig. 1. Flowchart showing patient enrollment.
CT = computed tomography, FOV = field of view, DICOM = Digital Imaging and Communications in Medicine.
CT acquisition
Cardiac CT scans were conducted at three hospitals using one of the following scanners: Revolution CT (GE Healthcare, Chicago, IL, USA), Somatom Definition (Siemens Healthineers, Forchheim, Germany), Somatom Force (Siemens Healthineers), or Aquilion ONE (Canon Medical Systems, Otawara, Japan). Each cardiac CT examination was comprised of one or more scans—coronary artery calcium (CAC) scoring, coronary computed tomography angiography (CCTA), and delayed phase imaging—depending on the purpose of each examination. At least one scan in each examination was reconstructed with a full FOV to encompass the entire lung field. The scanning range extended from the upper margin of the apex of the lung to the base of the heart or from below the carina to the base of the heart, depending on the indication. For pre-transcatheter aortic valve implantation evaluations, delayed phase imaging covered the area from the lower neck to the mid-thigh to include the thoracoabdominal region and common femoral artery.
The CAC scan was reconstructed with a 3-mm slice thickness and an increment of 3 mm or 1.5 mm, using a medium sharp kernel. The CCTA was reconstructed with a 0.75-mm slice thickness and an interval of 0.75 mm or 0.5 mm using a soft kernel. The delayed phase scan was reconstructed with a 0.63-mm slice thickness and an interval of 0.63 mm using a soft kernel. Slice thickness of ≤ 1.25 mm was defined as thin slice thickness, and that of ≥ 2.5 mm as thick slice thickness.
AI-CAD system
All cardiac CT images were analyzed to detect pulmonary nodules using a commercialized AI-CAD algorithm (AVIEW LCS, v1.1.46.15; Coreline Soft, Seoul, Korea).22,23,24,25 This AI-CAD was trained on the public LUNA16 dataset, which includes 888 CT scans with pulmonary nodules ranging in size from 3 mm to 30 mm.26 This algorithm was designed to achieve a sensitivity for lung nodule detection exceeding 0.8 at both the patient and nodule levels, a specificity greater than 0.6, an receiver operating characteristic area under the curve above 0.8, and a sensitivity exceeding 0.8 at a false-positive rate of fewer than two per scan.23 Among the various cardiac CT images, the CAC scan was prioritized for analysis. If a lung cancer lesion was not covered in the CAC scan, additional analyses were conducted on the CCTA and delayed phase scans. The AI-CAD system automatically recorded detected nodules, including data on their type, size, and lobar location. False-positive detections were not included in the analysis.
Radiology reports analysis
The initial radiology reports of all cardiac CT scans were retrospectively analyzed to include information about pulmonary nodules. All cardiac CT scans were interpreted by board-certified cardiothoracic radiologists from the three participating institutions. At the time of analysis, their experience as cardiothoracic radiologists ranged from 5 to 27 years at Hospital 1, 10 to 28 years at Hospital 2, and 10 years at Hospital 3. Based on the findings in radiology reports, a positive detection by the radiologist was defined as the presence of any description of the lesion that was subsequently diagnosed as lung cancer. This included not only statements indicating malignancy but also any descriptions of nodules or recommendations for further evaluation. Positive detection by combining the radiologist and AI-CAD was defined when the lesion was mentioned by the radiology report or identified by AI-CAD, given that we retrospectively analyzed existing radiology reports.
Additionally, data on the final coronary artery disease reporting and data system (RADS) category27 and the extent of coronary artery disease (defined as the number of coronary arteries with stenosis ≥ 50%) were extracted from radiology reports.
Characteristics and staging of lung cancer
All cardiac CT images were retrospectively reviewed by one of three expert cardiothoracic radiologists from each participating institution in the study. For each lung cancer lesion, the following characteristics were recorded: type (pure ground-glass, part-solid, or solid nodules), lobar location, centrality, size, lung-RADS category, the CT series in which the lung cancer lesion was identified, and the presumed stage of lung cancer at the time of CT examination and at the time of final diagnosis. Centrality was defined as central if the lesion was located within the FOV reconstructed to include the heart only and peripheral if it was located beyond the FOV of the heart. Lesion size was measured as the maximal long-axis diameter on one of three orthogonal planes. For part-solid nodules, the diameters of the total lesion and solid portion were measured. Lung cancer staging was determined following the American Joint Committee on Cancer 8th edition guidelines. All lesions were classified according to the presumed stage: limited (stages 1–2), locally advanced (stage 3), and advanced (stage 4). Delayed diagnosis was defined as cases in which lung cancer was diagnosed > 100 days after the cardiac CT scan.28
Statistical analysis
Categorical variables were presented as counts and percentages, whereas continuous variables were expressed as means with standard deviations or as medians with interquartile ranges, as appropriate. To compare groups based on the detection by radiologists or AI-CAD, the χ2 test or Fisher’s exact test was employed to assess differences in categorical variables, and the independent samples t-test or Mann–Whitney U test was utilized for continuous variables. The McNemar test was applied to compare detection rates between the two methods within the same group. For subgroup analysis, lesions were categorized into four groups based on detection by radiologists and AI-CAD. Subgroup analysis was conducted using analysis of variance for comparing continuous variables and the χ2 test or Fisher’s exact test for comparing categorical variables. To identify factors affecting the differences in AI-CAD detection rates across institutions, univariate and multivariate mixed-effects logistic regression analyses were conducted to predict the positivity of AI-CAD, considering institutions as a random effect and other variables as fixed effects. A P value of < 0.05 was considered statistically significant. All statistical analyses were conducted using R software (version 4.4.0; R Foundation for Statistical Computing, Vienna, Austria).
Ethics statement
This retrospective multicenter study was approved by the Institutional Review Boards of three participating hospitals (Severance Hospital, Yonsei University College of Medicine 4-2024-0798, Keimyung University Dongsan Hospital, Keimyung University College of Medicine 2024-04-053, and Seoul St. Mary’s Hospital, The Catholic University of Korea KC24RADI0779), with a waiver for informed consent.
RESULTS
Patient and lesion characteristics
Table 1 summarizes the characteristics of the patients and features of the lung cancer lesions included in this study. Most cardiac CT scans (comprising 53.8% of cases) were conducted owing to symptoms such as chest pain or dyspnea. Lung cancer lesions were most frequently located in the right upper lobe (29.6%), with a median lesion size of 19.8 mm. Lesions were predominantly solid (43.6%), followed by part-solid (36.4%) and pure ground-glass (20%) lesions. The Lung-RADS categories of the lesions were as follows: category 2 (20.5%), category 3 (13.5%), category 4A (14.1%), and category 4B (51.9%).
Table 1. Characteristics of patients (n = 581), CT scans (n = 652), and features of the lung cancer lesion (n = 665).
| Variables | Values | ||
|---|---|---|---|
| Clinical variables | |||
| Age, yr | 68.9 ± 9.3 | ||
| Sex (male) | 343 (59.0) | ||
| Family history of lung cancer | 42 (7.2) | ||
| Family history of coronary artery disease | 25 (4.3) | ||
| HTN | 380 (65.4) | ||
| DM | 174 (29.9) | ||
| Dyslipidemia | 94 (16.2) | ||
| COPD | 59 (10.2) | ||
| Smoking | |||
| Never smoker | 296 (50.9) | ||
| Former smoker | 237 (40.8) | ||
| Current smoker | 48 (8.3) | ||
| Pack-year | 35.0 (20.0–50.0) | ||
| History of lung cancer | 7 (1.2) | ||
| Indications for cardiac CT | |||
| Screening | 46 (7.1) | ||
| Symptomatic | 351 (53.8) | ||
| Others | 255 (39.1) | ||
| Lesion characteristics | |||
| Location | |||
| Right upper lobe | 197 (29.6) | ||
| Right middle lobe | 59 (8.9) | ||
| Right lower lobe | 132 (19.8) | ||
| Left upper lobe | 172 (25.9) | ||
| Left lower lobe | 105 (15.8) | ||
| Total size, mm | 19.8 (13.2–28.4) | ||
| Type | |||
| Solid | 290 (43.6) | ||
| Part-solid | 242 (36.4) | ||
| Pure ground glass | 133 (20.0) | ||
| Lung-RADS category | |||
| 2 | 136 (20.5) | ||
| 3 | 90 (13.5) | ||
| 4A | 94 (14.1) | ||
| 4B | 345 (51.9) | ||
| Histologic type | |||
| Adenocarcinoma | 478 (71.9) | ||
| Squamous cell carcinoma | 61 (9.2) | ||
| Small cell carcinoma | 32 (4.8) | ||
| Large cell carcinoma | 4 (0.6) | ||
| Non-small cell carcinoma, not otherwise specified | 4 (0.6) | ||
| Poorly differentiated adenocarcinoma | 4 (0.6) | ||
| Neuroendocrine tumor | 1 (0.1) | ||
| Unknown (clinical diagnosis) | 81 (12.2) | ||
The data are presented as the mean ± standard deviation, median (interquartile range), or number (%).
CT = computed tomography, HTN = hypertension, DM = diabetes mellitus, COPD = chronic obstructive pulmonary disease, RADS = reporting and data system.
Among the 665 lung cancer cases, 584 (87.8%) were pathologically confirmed, with adenocarcinomas being the most prevalent histologic type (478 of 584 cases, 81.8%). The remaining 81 cases (12.2%) were clinically diagnosed without histologic confirmation.
Detection rates for lung cancer lesions by radiologists and AI-CAD
Table 2 shows the overall detection rates of lung cancer lesions by radiologists alone, AI-CAD alone, and the combination of both. The radiologist achieved a detection rate of 76.2%, while AI-CAD detected 77.4%, with no significant difference (P = 0.551) between the two methods. However, when combining the radiologist and AI-CAD, the detection rate increased significantly to 90.4%, compared to that of the radiologist alone (P < 0.001).
Table 2. Detection rates for overall lung nodules and lung nodules based on the lung cancer stage of the radiologist, AI-CAD, and the combination of the radiologist and AI-CAD.
| Variables | Radiologist | CAD | Radiologist + AI-CADa | P value (radiologist vs. AI-CAD) | P value (radiologist vs. radiologist + AI-CAD) | |
|---|---|---|---|---|---|---|
| Total (N = 665) | 507/665 (76.2) | 515/665 (77.4) | 601/665 (90.4) | 0.551 | < 0.001 | |
| Hospital 1 (n = 583) | 445/583 (76.3) | 458/583 (78.6) | 530/583 (90.9) | 0.300 | < 0.001 | |
| Hospital 2 (n = 47) | 31/47 (66.0) | 28/47 (59.6) | 37/47 (78.7) | 0.439 | 0.014 | |
| Hospital 3 (n = 35) | 31/35 (88.6) | 29/35 (82.9) | 34/35 (97.1) | 0.479 | 0.083 | |
| Presumed lung cancer stage at cardiac CT | ||||||
| Limited (n = 613) | 458/613 (74.7) | 492/613 (80.3) | 551/613 (89.9) | 0.006 | < 0.001 | |
| Locally advanced (n = 25) | 22/25 (88.0) | 12/25 (48.0) | 23/25 (92.0) | 0.004 | 0.317 | |
| Advanced (n = 27) | 27/27 (100) | 11/27 (40.7) | 27/27 (100) | < 0.001 | > 0.999 | |
AI = artificial intelligence, CAD = computer-aided detection, CT = computed tomography.
aThe lesion is mentioned through a radiology report or identified via AI-CAD.
When analyzed across institutions, the detection rates of the radiologist were 76.3%, 66.0%, and 88.6% for Hospitals 1, 2, and 3, respectively. The corresponding detection rates for AI-CAD were 78.6%, 59.6%, and 82.9%, respectively. Combining the radiologist and AI-CAD resulted in significantly improved detection rates of 90.9%, 78.7%, and 97.1%, respectively. This improvement was statistically significant for Hospitals 1 and 2 (P < 0.001 and P = 0.014, respectively) but not for Hospital 3 (P = 0.083).
AI-CAD exhibited a higher detection rate for limited-stage cancer than that of radiologists (80.3% vs. 74.7%, P = 0.006). Combining the radiologist and AI-CAD significantly improved detection to 89.9% (P < 0.001). For locally advanced cancer, the radiologist exhibited a significantly higher detection rate than that of AI-CAD (88.0% vs. 48.0%, P = 0.004). In advanced-stage cancers, the radiologist achieved a 100% detection rate, whereas AI-CAD detected 40.7% (P < 0.001).
No significant difference in detection rates was observed between cases with single-vessel (75.9%, 441 of 581) and multiple-vessel coronary artery disease (78.6%, 66 of 84) (P = 0.689).
Characteristics of nodules detected and missed by radiologists and AI-CAD
Table 3 shows the comparison of the characteristics of nodules detected and missed by the radiologist. The average size of nodules detected by the radiologist was significantly larger (25.8 mm) than that of undetected nodules (15.8 mm) (P < 0.001). Solid nodules were more frequently detected, whereas pure ground-glass nodules were more frequently missed (P = 0.002). Other factors, such as lesion location, CT characteristics, or institution, did not significantly influence detection rates.
Table 3. Comparison of the characteristics between nodules detected and those missed by the radiologist.
| Variables | Detect (n = 507) | Miss (n = 158) | P value | |
|---|---|---|---|---|
| Total size, mm | 25.8 ± 16.4 | 15.8 ± 11.8 | < 0.001 | |
| Lesion type | 0.002 | |||
| Solid | 236 (81.4) | 54 (18.6) | ||
| Part-solid | 185 (76.4) | 57 (23.6) | ||
| Pure ground glass | 86 (64.7) | 47 (35.3) | ||
| Location | 0.146 | |||
| Upper | 278 (75.7) | 89 (24.3) | ||
| Middle | 51 (86.4) | 8 (13.6) | ||
| Lower | 178 (74.5) | 61 (25.5) | ||
| Centrality | 0.580 | |||
| Peripheral | 332 (77.0) | 99 (23.0) | ||
| Central | 175 (74.8) | 59 (25.2) | ||
| Hospital | 0.058 | |||
| Hospital 1 | 445 (76.3) | 138 (23.7) | ||
| Hospital 2 | 31 (66.0) | 16 (34.0) | ||
| Hospital 3 | 31 (88.6) | 4 (11.4) | ||
| CT protocol | 0.451 | |||
| CAC scoring | 10 (62.5) | 6 (37.5) | ||
| Coronary artery evaluation | 371 (75.9) | 118 (24.1) | ||
| pre-RFA | 78 (80.4) | 19 (19.6) | ||
| Valve, pre-TAVI | 48 (76.2) | 15 (23.8) | ||
| CT series | 0.669 | |||
| Calcium score | 460 (76.5) | 141 (23.5) | ||
| CCTA | 10 (66.7) | 5 (33.3) | ||
| Delayed | 37 (75.5) | 12 (24.5) | ||
| CT slice thickness | 0.426 | |||
| Thin (≤ 1.25 mm) | 42 (71.2) | 17 (28.8) | ||
| Thick (≥ 2.5 mm) | 465 (76.7) | 141 (23.3) | ||
| CT scan range (Z axis) | 0.601 | |||
| Whole thorax | 405 (77.3) | 119 (22.7) | ||
| Heart | 67 (71.3) | 27 (28.7) | ||
| Upper thorax | 24 (72.7) | 9 (27.3) | ||
| Thoraco-abdomen | 11 (78.6) | 3 (21.4) | ||
The data are presented as the mean ± standard deviation or number (%).
CT = computed tomography, CAC = coronary artery calcium, RFA = radiofrequency ablation, TAVI = transcatheter aortic valve implantation, CCTA = coronary computed tomography angiography.
Table 4 shows the comparison of the characteristics of nodules detected and missed by the AI-CAD. The average total size of detected nodules was 21.7 mm, smaller than that of undetected nodules (29.3 mm) (P < 0.001). AI-CAD exhibited a higher detection rate for part-solid and pure ground-glass nodules, but it missed a greater proportion of solid nodules (P < 0.001). Additionally, AI-CAD exhibited superior performance in detecting nodules located in peripheral lung regions compared to those in central locations (P < 0.001).
Table 4. Comparison of the characteristics between nodules detected and those missed by the artificial intelligence-based computer-aided detection.
| Variables | Detect (n = 515) | Miss (n = 150) | P value | |
|---|---|---|---|---|
| Total size, mm | 21.669 ± 13.254 | 29.317 ± 22.200 | < 0.001 | |
| Lesion type | < 0.001 | |||
| Solid | 199 (68.6) | 91 (31.4) | ||
| Part-solid | 213 (88.0) | 29 (12.0) | ||
| Pure ground glass | 103 (77.4) | 30 (22.6) | ||
| Location | 0.064 | |||
| Upper | 294 (80.1) | 73 (19.9) | ||
| Middle | 48 (81.4) | 11 (18.6) | ||
| Lower | 173 (72.4) | 66 (27.6) | ||
| Centrality | <0.001 | |||
| Peripheral | 354 (82.1) | 77 (17.9) | ||
| Central | 161 (68.8) | 73 (31.2) | ||
| Hospital | 0.008 | |||
| Hospital 1 | 458 (78.6) | 125 (21.4) | ||
| Hospital 2 | 28 (59.6) | 19 (40.4) | ||
| Hospital 3 | 29 (82.9) | 6 (17.1) | ||
| CT protocol | 0.580 | |||
| CAC scoring | 12 (75.0) | 4 (25.0) | ||
| Coronary artery evaluation | 377 (77.1) | 112 (22.9) | ||
| Pre-RFA | 73 (75.3) | 24 (24.7) | ||
| Valve, pre-TAVI | 53 (84.1) | 10 (15.9) | ||
| CT series | 0.971 | |||
| Calcium score | 465 (77.4) | 136 (22.6) | ||
| CCTA | 12 (80.0) | 3 (20.0) | ||
| Delayed | 38 (77.6) | 11 (22.4) | ||
| CT slice thickness | > 0.999 | |||
| Thin (≤ 1.25 mm) | 46 (78.0) | 13 (22.0) | ||
| Thick (≥ 2.5 mm) | 469 (77.4) | 137 (22.6) | ||
| CT scan range (Z axis) | 0.997 | |||
| Whole thorax | 405 (77.3) | 119 (22.7) | ||
| Heart | 73 (77.7) | 21 (22.3) | ||
| Upper thorax | 26 (78.8) | 7 (21.2) | ||
| Thoraco-abdomen | 11 (78.6) | 3 (21.4) | ||
The data are presented as the mean ± standard deviation or number (%).
CT = computed tomography, CAC = coronary artery calcium, RFA = radiofrequency ablation, TAVI = transcatheter aortic valve implantation, CCTA = coronary computed tomography angiography.
Analysis of lesion and CT scan characteristics in detection subgroups
The subgroup analysis showed significant differences in lesion total size, type, and centrality among the four subgroups (Table 5). Figs. 2, 3, 4 present representative CT images of lung cancer lesions for each subgroup. In lesion total size, nodules detected by the radiologist but missed by AI-CAD (Group 2) were significantly larger than those in other subgroups (P < 0.001). Regarding the type, nodules in Group 2 were predominantly solid, whereas those detected only by AI-CAD (Group 3) were predominantly part-solid nodules (P < 0.001).
Table 5. Analysis of lesion and CT scan characteristics in subgroups based on detection status by the radiologist and AI-CAD.
| Variables | Subgroup | P value (for four groups) | P value (for group 3 vs. 4) | ||||
|---|---|---|---|---|---|---|---|
| Group 1: both detect (n = 421) | Group 2: radiologist detect + AI-CAD miss (n = 86) | Group 3: radiologist miss + AI-CAD detect (n = 94) | Group 4: both miss (n = 64) | ||||
| Total size, mm | 23.086 ± 13.812 | 38.873 ± 21.304 | 15.322 ± 7.7137 | 16.477 ± 16.159 | < 0.001 | 0.549 | |
| Lesion type | < 0.001 | 0.024 | |||||
| Solid | 171 (40.6) | 65 (75.6) | 28 (29.8) | 26 (40.6) | |||
| Part-solid | 171 (40.6) | 14 (16.3) | 42 (44.7) | 15 (23.4) | |||
| Pure ground glass | 79 (18.8) | 7 (8.1) | 24 (25.5) | 23 (36.0) | |||
| Location | 0.099 | 0.132 | |||||
| Upper | 239 (56.8) | 39 (45.3) | 55 (58.5) | 34 (53.1) | |||
| Middle | 41 (9.7) | 10 (11.6) | 7 (7.4) | 1 (1.6) | |||
| Lower | 141 (33.5) | 37 (43.0) | 32 (34.1) | 29 (45.3) | |||
| Centrality | 0.001 | 0.123 | |||||
| Peripheral | 290 (68.9) | 42 (48.8) | 64 (68.1) | 35 (54.7) | |||
| Central | 131 (31.1) | 44 (51.2) | 30 (31.9) | 29 (45.3) | |||
| Hospital | 0.033 | 0.145 | |||||
| Hospital 1 | 373 (88.6) | 72 (83.7) | 85 (90.4) | 53 (82.8) | |||
| Hospital 2 | 22 (5.2) | 9 (10.5) | 6 (6.4) | 10 (15.6) | |||
| Hospital 3 | 26 (6.2) | 5 (5.8) | 3 (3.2) | 1 (1.6) | |||
| CT protocol | 0.199 | 0.506 | |||||
| CAC scoring | 10 (2.4) | 0 (0) | 2 (2.1) | 4 (6.2) | |||
| Coronary artery evaluation | 307 (72.9) | 64 (74.4) | 70 (74.5) | 48 (75.0) | |||
| Pre-RFA | 60 (14.3) | 18 (20.9) | 13 (13.8) | 6 (9.4) | |||
| Valve, Pre-TAVI | 44 (10.4) | 4 (4.7) | 9 (9.6) | 6 (9.4) | |||
| CT series | 0.922 | 0.784 | |||||
| Calcium score | 380 (90.3) | 80 (93.0) | 85 (90.4) | 56 (87.5) | |||
| CCTA | 9 (2.1) | 1 (1.2) | 3 (3.2) | 2 (3.1) | |||
| Delayed | 32 (7.6) | 5 (5.8) | 6 (6.4) | 6 (9.4) | |||
| CT slice thickness | 0.552 | 0.748 | |||||
| Thin (≤ 1.25 mm) | 37 (8.8) | 5 (5.8) | 9 (9.6) | 8 (12.5) | |||
| Thick (≥ 2.5 mm) | 384 (91.2) | 81 (94.2) | 85 (90.4) | 56 (87.5) | |||
| CT scan range (Z axis) | 0.853 | 0.740 | |||||
| Whole thorax | 332 (78.8) | 73 (84.9) | 73 (77.7) | 46 (71.9) | |||
| Heart | 58 (13.8) | 9 (10.5) | 15 (16.0) | 12 (18.8) | |||
| Upper thorax | 21 (5.0) | 3 (3.5) | 5 (5.3) | 4 (6.2) | |||
| Thoraco-abdomen | 10 (2.4) | 1 (1.2) | 1 (1.0) | 2 (3.1) | |||
The data are presented as the mean ± standard deviation or number (%).
CT = computed tomography, AI-CAD = artificial intelligence-based computer-aided detection, CAC = coronary artery calcium, RFA = radiofrequency ablation, TAVI = transcatheter aortic valve implantation, CCTA = coronary computed tomography angiography.
Fig. 2. Representative CT images of lung cancers detected by the radiologist. (A) A 53-year-old male patient undergoing cardiac CT for coronary artery bypass graft follow-up, revealing a 22 mm solid nodule (arrow). The radiologist reported the possibility of malignancy, and a transbronchial lung biopsy conducted 28 days later confirmed squamous cell carcinoma. AI-CAD also identifies the lesion. (B) A 61-year-old male patient undergoing cardiac CT as part of a health checkup, revealing a 61 mm solid mass (arrow). The radiologist reported the possibility of malignancy, and a lobectomy conducted 21 days later confirmed adenocarcinoma. AI-CAD failed to detect the lesion.
CT = computed tomography, AI-CAD = artificial intelligence-based computer-aided detection.
Fig. 3. Representative CT images of lung cancer detected by AI-CAD but missed by the radiologist. (A) A 65-year-old male patient undergoing cardiac CT for chest pain, revealing a 15 mm part-solid nodule with a 4 mm solid component (arrow). AI-CAD identifies the lesion (the yellow dotted line indicates the non-solid component; the red dotted line indicates the solid component), but at the time of the CT, the radiologist did not detect the lesion. (B) Three years later, chest CT revealed a 50 mm mass (arrow) at the same location. Biopsy confirms adenocarcinoma.
CT = computed tomography, AI-CAD = artificial intelligence-based computer-aided detection.
Fig. 4. Representative CT images of lung cancer missed by the radiologist and AI-CAD. (A) A 73-year-old male patient undergoing cardiac CT for electrocardiogram abnormalities, revealing a 20 mm pure ground-glass nodule (arrow). At the time of the CT, the radiologist failed to identify the lesion, and retrospective analysis confirms that AI-CAD also failed to detect it. (B) Two years later, chest CT reveals the development of an 8 mm solid component within the nodule (arrow). After lobectomy, the lesion is confirmed as adenocarcinoma.
CT = computed tomography, AI-CAD = artificial intelligence-based computer-aided detection.
When comparing nodules missed by the radiologist but detected by AI-CAD (Group 3) with those missed by the radiologist and AI-CAD (Group 4), nodules in Group 3 exhibited a significantly higher proportion of part-solid nodules (P = 0.024). Other variables showed no statistically significant difference between the two subgroups.
For the 64 lesions (Group 4) that were missed by both the radiologists and the AI-CAD system, a comprehensive review was conducted by an expert cardiothoracic radiologist. The findings revealed the following characteristics: 19 faint nonsolid nodules, 16 nodules located in the subpleural region, nine tiny nodules measuring less than 6 mm, six nodules adjacent to vessels, five nodules mimicking tuberculous sequelae, two nodules resembling pneumonic consolidation, two perifissural nodules, two nodules located at the margins of the scan range, one nodule adjacent to a cyst, one nodule associated with diffuse pneumoconiosis, and one lesion mimicking collapse.
Factors affecting variability in detection rate among institutions
Supplementary Table 1 shows the differences in lesion and CT scan characteristics across the three institutions. Significant differences were observed among the three institutions regarding lesion type (P < 0.001), location (P = 0.003), CT protocol (P < 0.001), CT series (P < 0.001), CT slice thickness (P = 0.008), and CT scan range (P < 0.001).
Univariate analysis identified total lesion size (P < 0.001), lesion type (P < 0.001), lower lobar location (P = 0.028), and centrality (P < 0.001) as being significantly associated with detection rates of AI-CAD (Supplementary Table 2). Multivariate analysis identified smaller total lesion size (P < 0.001), part-solid type (P < 0.001), and peripheral location (P = 0.017) as independent factors for predicting positive AI-CAD detection.
Diagnostic delays owing to missed nodule detections
Among the 158 patients whose nodules were not reported by the radiologist, 94.2% (149 of 158) experienced a delay of > 100 days from the initial CT to the confirmation of lung cancer diagnosis. The average time to diagnosis for these patients was 1,538.1 days. In these cases, AI-CAD achieved a detection rate of 58.5% (87 of 149).
Among the 149 patients with diagnostic delays, 74% (110 of 149) had an upgrade in the stage of lung cancer between the time of the initial CT scan and the final diagnosis. The AI-CAD identified lesions in 60.7% (67 of 110) of these cases that experienced such an upstaging.
DISCUSSION
The findings of our multicenter study demonstrate the potential of the AI-CAD system to improve the performance of radiologists in lung cancer detection rates on cardiac CT. Although the AI-CAD alone achieved detection rates (77.4%) of lung cancer comparable to those of radiologists (76.2%), the combination of AI-CAD and radiologist interpretation led to significantly enhanced detection rates (90.4%). This indicates a complementary role of AI in clinical settings. AI-CAD demonstrates a distinct advantage in identifying early-stage lung cancers, particularly small, part-solid nodules that are frequently missed by radiologists. Most cases of lung cancer missed in the initial radiology reports were observed to have progressed to more advanced stages owing to delayed diagnosis. However, AI-CAD successfully identified a substantial portion of these missed cases.
Recent advancements in AI technology show that commercial AI-based software has been developed to detect nodules on chest CT scans. Several previous studies show that AI assistance in detecting nodules on chest CT scans enhances the sensitivity of radiologists.16,17,18,19,20,21,29 These studies show lower sensitivity (25.2–64%) for radiologists than that observed in ours, probably because they included residents in the interpretation process, whereas our study involved only board-certified cardiothoracic radiologists. Additionally, the sensitivity of radiologists with AI assistance in previous studies (up to 99.1%) is much higher than that observed in ours. We attribute this discrepancy in sensitivity to the different characteristics between our study and that of previous ones. First, while previous studies analyzed all lung nodules, our study focused specifically on lung cancer. Additionally, we utilized existing reports for analysis, whereas reader studies were employed in previous studies. Last, previous studies employed chest CT protocols, which are more suitable for detecting lung nodules. Although direct comparison with our study is challenging owing to differences in study settings, the direction of our findings aligns with the conclusion that AI enhances the sensitivity of radiologists.
In our study, AI-CAD demonstrated a strong ability to detect small, peripheral part-solid nodules and was particularly effective in identifying early-stage lung cancers compared to radiologists. Early detection of lung cancer is well-documented to improve patient outcomes,30 especially when the disease is identified at an early stage.31,32 However, radiologists interpreting cardiac CT scans are primarily focused on diagnosing cardiac pathology, which may lead to incomplete detection of extracardiac findings.11,33 In our study, small and subsolid lung cancers were more frequently missed by radiologists but were detected by AI-CAD. Conversely, advanced lung cancers presenting as large solid masses, which were conspicuous and readily identified by radiologists, were occasionally not detected by AI-CAD. These differences in lesion characteristics highlight the complementary strengths of AI-CAD and radiologists. Integrating AI-CAD into cardiac CT workflows can complement the interpretation of radiologists by improving early detection of incidental lung cancer. This can ultimately prevent diagnostic delays and progression of the lung cancer stage and improve patient outcomes.
With the increasing number of chest CT examinations and the growing detection of pulmonary nodules, concerns regarding overdiagnosis have been rising. High sensitivity for small subsolid nodules may result in a greater number of invasive procedures and additional downstream imaging studies. However, adherence to appropriate follow-up protocols, such as those recommended by the Fleischner Society guidelines, may help minimize both overdiagnosis and delayed diagnosis, thereby maximizing the benefits of screening through follow-up examinations performed at appropriate intervals.34,35
In our study cohort, 50.9% of patients were non-smokers and therefore would not have been eligible for current lung cancer screening programs.28 Our findings suggest that detecting lung cancer on CT scans performed for non-screening purposes in such individuals may help prevent disease progression by identifying malignancies at an earlier stage. These results indicate that the benefits of lung cancer screening may also extend to populations not traditionally targeted by current screening guidelines. Further well-designed studies are warranted to validate this potential benefit.
The CT scan parameters and extent of coronary artery disease did not significantly influence the detection rate of radiologists. Although thinner slice thickness is conventionally associated with an increased detection rate of lung nodules,36 most of our study population had lung cancer nodules seen on thick slice scans. As tiny nodules < 6 mm in diameter do not typically warrant follow-up or intervention,37,38 it may not be necessary to use thin slice thickness specifically for detecting lung lesions on cardiac CT. Additionally, while a broader scan range could theoretically reduce the attention of radiologists from the lungs to inspecting other organs within the FOV, we observed no significant effect of scan range on detection rates in our study. Similarly, a greater extent of coronary artery disease, which could potentially distract radiologists from lung findings, exhibited no measurable effect on their detection rate. These findings suggest that the detection rate of radiologists is primarily affected by the intrinsic characteristics of the lesions, such as their type and size. The limitations in detecting small, part-solid nodules can be compensated by AI-CAD assistance.
In our study, the detection rate of AI-CAD varied among institutions despite all three institutions utilizing the same version of the AI-CAD system. This variability in the performance of AI-CAD can be attributed to variations in lesion characteristics and CT scan parameters across institutions. Key independent factors influencing AI-CAD detection include total lesion size, part-solid type, and lesion centrality. Hospital 3 specifically exhibited larger total lesion sizes and a lower proportion of part-solid lesions compared to those in other institutions. These differences probably account for the lack of significant improvement in detection rates when combining AI-CAD with radiologist interpretations. Concerns regarding overfitting and limited generalizability of AI algorithms remain prominent in this field,39 as differences in basic dataset characteristics were observed to influence the AI model performance. Validating AI models on heterogeneous data from multiple institutions remains a significant challenge, and our findings highlight the importance of addressing this issue.
Our study has some limitations. First, its retrospective design relied on previously generated reports. Consequently, the AI-CAD–assisted detection rate for radiologists was inferred indirectly. In addition, reviewing existing reports makes it difficult to differentiate truly missed lesions from those that were intentionally unreported because of limited clinical relevance. Second, the study cohort consisted exclusively of patients with lung cancer, and no control group was included. Consequently, analyses of specificity, false-positive rate, and positive predictive value could not be performed. In addition, we excluded false-positive detections by the AI-CAD from our analysis. In a real-world workflow, false-positive detections generated by AI-CAD may decrease reading efficiency and potentially lead to overdiagnosis. A future reader study is warranted to assess specificity and evaluate the real-world impact of AI-CAD implementation on radiologists’ workflow, including reading time, in the context of cardiac CT.16,29 Third, some nodules categorized as diagnostic delays may have represented precancerous lesions at the time of cardiac CT, which could not be verified in our study. Nevertheless, the implementation of AI-CAD may facilitate the detection of pulmonary nodules and enable appropriate follow-up strategies, thereby allowing earlier diagnosis if these lesions progress to lung cancer. Fourth, this study reported results at the lesion level, which may have overstated the utility of the AI-CAD system. Fifth, our study was limited to only cardiac CT scans conducted with a wide FOV. Given the debate regarding the appropriate FOV for cardiac CT scans,40 further studies are necessary to determine whether AI-CAD remains effective for scans with an FOV limited to the heart. Finally, we did not explore whether the introduction of AI-CAD could improve patient survival rate in our study. However, our findings show that a significant proportion of lung cancers that were missed by the radiologist, leading to diagnostic delays and stage upgrades, could be identified with AI-CAD.
In conclusion, AI-CAD demonstrated the potential to improve the detection rate of incidental lung cancer by identifying a subset of lesions that were initially overlooked by radiologists on cardiac CT. The AI-CAD system was particularly effective in identifying early-stage lung cancers presenting small subsolid nodules, and its application has the potential to reduce diagnostic delays and prevent stage progression in lung cancer.
Footnotes
Funding: This study was supported by a faculty research grant of Yonsei University College of Medicine (grant number: 6-2024-0097).
Disclosure: The authors have no potential conflicts of interest to disclose.
- Conceptualization: Suh YJ, Chang S.
- Data curation: Kim JY.
- Formal analysis: Son J, Kim JY.
- Funding acquisition: Suh YJ.
- Investigation: Son J, Kim JY.
- Methodology: Suh YJ, Chang S.
- Validation: Son J, Kim JY.
- Writing - original draft: Son J.
- Writing - review & editing: Kim JY, Chang S, Suh YJ.
SUPPLEMENTARY MATERIALS
Comparison of lesion and CT scan characteristics by institutions
Univariate and multivariate mixed-effects logistic regression analyses for variables affecting differences in AI-CAD detection rate across institutions
References
- 1.Gulati M, Levy PD, Mukherjee D, Amsterdam E, Bhatt DL, Birtcher KK, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR guideline for the evaluation and diagnosis of chest pain: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Cardiovasc Comput Tomogr. 2022;16(1):54–122. doi: 10.1016/j.jcct.2021.11.009. [DOI] [PubMed] [Google Scholar]
- 2.Shaw LJ, Blankstein R, Bax JJ, Ferencik M, Bittencourt MS, Min JK, et al. Society of Cardiovascular Computed Tomography/North American Society of Cardiovascular Imaging–expert consensus document on coronary CT imaging of atherosclerotic plaque. J Cardiovasc Comput Tomogr. 2021;15(2):93–109. doi: 10.1016/j.jcct.2020.11.002. [DOI] [PubMed] [Google Scholar]
- 3.Carminati M, Agnifili M, Arcidiacono C, Brambilla N, Bussadori C, Butera G, et al. Role of imaging in interventions on structural heart disease. Expert Rev Cardiovasc Ther. 2013;11(12):1659–1676. doi: 10.1586/14779072.2013.854166. [DOI] [PubMed] [Google Scholar]
- 4.Blanke P, Weir-McCall JR, Achenbach S, Delgado V, Hausleiter J, Jilaihawi H, et al. Computed tomography imaging in the context of transcatheter aortic valve implantation (TAVI)/transcatheter aortic valve replacement (TAVR) an expert consensus document of the Society of Cardiovascular Computed Tomography. JACC Cardiovasc Imaging. 2019;12(1):1–24. doi: 10.1016/j.jcmg.2018.12.003. [DOI] [PubMed] [Google Scholar]
- 5.Lu MT, Onuma OK, Massaro JM, D’Agostino RB, Sr, O’Donnell CJ, Hoffmann U. Lung cancer screening eligibility in the community: cardiovascular risk factors, coronary artery calcification, and cardiovascular events. Circulation. 2016;134(12):897–899. doi: 10.1161/CIRCULATIONAHA.116.023957. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Koene RJ, Prizment AE, Blaes A, Konety SH. Shared risk factors in cardiovascular disease and cancer. Circulation. 2016;133(11):1104–1114. doi: 10.1161/CIRCULATIONAHA.115.020406. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Johnson KM, Dennis JM, Dowe DA. Extracardiac findings on coronary CT angiograms: limited versus complete image review. AJR Am J Roentgenol. 2010;195(1):143–148. doi: 10.2214/AJR.08.1050. [DOI] [PubMed] [Google Scholar]
- 8.Earls JP. The pros and cons of searching for extracardiac findings at cardiac CT: studies should be reconstructed in the maximum field of view and adequately reviewed to detect pathologic findings. Radiology. 2011;261(2):342–346. doi: 10.1148/radiol.11111099. [DOI] [PubMed] [Google Scholar]
- 9.Flor N, Di Leo G, Squarza SAC, Tresoldi S, Rulli E, Cornalba G, et al. Malignant incidental extracardiac findings on cardiac CT: systematic review and meta-analysis. AJR Am J Roentgenol. 2013;201(3):555–564. doi: 10.2214/AJR.12.10306. [DOI] [PubMed] [Google Scholar]
- 10.Kim TJ, Han DH, Jin KN, Lee KW. Lung cancer detected at cardiac CT: prevalence, clinicoradiologic features, and importance of full-field-of-view images. Radiology. 2010;255(2):369–376. doi: 10.1148/radiol.10091083. [DOI] [PubMed] [Google Scholar]
- 11.Kim JY, Suh YJ, Nam K, Choi BW. Lung cancer detected on coronary artery calcium scoring computed tomography: factors delaying diagnosis and predictors of survival. Acta Radiol. 2019;60(9):1118–1126. doi: 10.1177/0284185118815297. [DOI] [PubMed] [Google Scholar]
- 12.Gu Y, Chi J, Liu J, Yang L, Zhang B, Yu D, et al. A survey of computer-aided diagnosis of lung nodules from CT scans using deep learning. Comput Biol Med. 2021;137:104806. doi: 10.1016/j.compbiomed.2021.104806. [DOI] [PubMed] [Google Scholar]
- 13.Cui X, Zheng S, Heuvelmans MA, Du Y, Sidorenkov G, Fan S, et al. Performance of a deep learning-based lung nodule detection system as an alternative reader in a Chinese lung cancer screening program. Eur J Radiol. 2022;146:110068. doi: 10.1016/j.ejrad.2021.110068. [DOI] [PubMed] [Google Scholar]
- 14.Cui S, Ming S, Lin Y, Chen F, Shen Q, Li H, et al. Development and clinical application of deep learning model for lung nodules screening on CT images. Sci Rep. 2020;10(1):13657. doi: 10.1038/s41598-020-70629-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Li R, Xiao C, Huang Y, Hassan H, Huang B. Deep learning applications in computed tomography images for pulmonary nodule detection and diagnosis: a review. Diagnostics (Basel) 2022;12(2):298. doi: 10.3390/diagnostics12020298. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Hsu HH, Ko KH, Chou YC, Wu YC, Chiu SH, Chang CK, et al. Performance and reading time of lung nodule identification on multidetector CT with or without an artificial intelligence-powered computer-aided detection system. Clin Radiol. 2021;76(8):626.e23–e32. doi: 10.1016/j.crad.2021.04.006. [DOI] [PubMed] [Google Scholar]
- 17.Zhang Y, Jiang B, Zhang L, Greuter MJW, de Bock GH, Zhang H, et al. Lung nodule detectability of artificial intelligence-assisted CT image reading in lung cancer screening. Curr Med Imaging Rev. 2022;18(3):327–334. doi: 10.2174/1573405617666210806125953. [DOI] [PubMed] [Google Scholar]
- 18.Lan CC, Hsieh MS, Hsiao JK, Wu CW, Yang HH, Chen Y, et al. Deep learning-based artificial intelligence improves accuracy of error-prone lung nodules. Int J Med Sci. 2022;19(3):490–498. doi: 10.7150/ijms.69400. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Murchison JT, Ritchie G, Senyszak D, Nijwening JH, van Veenendaal G, Wakkie J, et al. Validation of a deep learning computer aided system for CT based lung nodule detection, classification, and growth rate estimation in a routine clinical population. PLoS One. 2022;17(5):e0266799. doi: 10.1371/journal.pone.0266799. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Hu Q, Wang S, Chen C, Kang S, Sun Z, Wang Y, et al. Comparison of two reader modes of computer-aided diagnosis in lung nodules on low-dose chest CT scan. J Innov Opt Health Sci. 2022;15(2):2250013 [Google Scholar]
- 21.Sun K, Chen S, Zhao J, Wang B, Yang Y, Wang Y, et al. Convolutional neural network-based diagnostic model for a solid, indeterminate solitary pulmonary nodule or mass on computed tomography. Front Oncol. 2021;11:792062. doi: 10.3389/fonc.2021.792062. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Kim H, Kim HY, Goo JM, Kim Y. Lung cancer CT screening and lung-RADS in a tuberculosis-endemic country: the Korean Lung Cancer Screening Project (K-LUCAS) Radiology. 2020;296(1):181–188. doi: 10.1148/radiol.2020192283. [DOI] [PubMed] [Google Scholar]
- 23.Kondrashova R, Klimeš F, Kaireit TF, May K, Barkhausen J, Stiebeler S, et al. Comparison of AI software tools for automated detection, quantification and categorization of pulmonary nodules in the HANSE LCS trial. Sci Rep. 2024;14(1):27809. doi: 10.1038/s41598-024-78568-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Lee JJB, Suh YJ, Oh C, Lee BM, Kim JS, Chang Y, et al. Automated computer-aided detection of lung nodules in metastatic colorectal cancer patients for the identification of pulmonary oligometastatic disease. Int J Radiat Oncol Biol Phys. 2022;114(5):1045–1052. doi: 10.1016/j.ijrobp.2022.08.042. [DOI] [PubMed] [Google Scholar]
- 25.Lancaster HL, Zheng S, Aleshina OO, Yu D, Yu Chernina V, Heuvelmans MA, et al. Outstanding negative prediction performance of solid pulmonary nodule volume AI for ultra-LDCT baseline lung cancer screening risk stratification. Lung Cancer. 2022;165:133–140. doi: 10.1016/j.lungcan.2022.01.002. [DOI] [PubMed] [Google Scholar]
- 26.Armato SG, 3rd, McLennan G, Bidaut L, McNitt-Gray MF, Meyer CR, Reeves AP, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans. Med Phys. 2011;38(2):915–931. doi: 10.1118/1.3528204. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Cury RC, Leipsic J, Abbara S, Achenbach S, Berman D, Bittencourt M, et al. CAD-RADS™ 2.0–2022 coronary artery disease-reporting and data system: an expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Cardiology (ACC), the American College of Radiology (ACR), and the North America Society of Cardiovascular Imaging (NASCI) JACC Cardiovasc Imaging. 2022;15(11):1974–2001. doi: 10.1016/j.jcmg.2022.07.002. [DOI] [PubMed] [Google Scholar]
- 28.Yankelevitz DF, Yip R, Henschke CI. Impact on prognosis of stage I non-small cell lung cancer secondary to delays in diagnostic workup. Radiology. 2024;313(1):e240420. doi: 10.1148/radiol.240420. [DOI] [PubMed] [Google Scholar]
- 29.Kozuka T, Matsukubo Y, Kadoba T, Oda T, Suzuki A, Hyodo T, et al. Efficiency of a computer-aided diagnosis (CAD) system with deep learning in detection of pulmonary nodules on 1-mm-thick images of computed tomography. Jpn J Radiol. 2020;38(11):1052–1061. doi: 10.1007/s11604-020-01009-0. [DOI] [PubMed] [Google Scholar]
- 30.de Koning HJ, van der Aalst CM, de Jong PA, Scholten ET, Nackaerts K, Heuvelmans MA, et al. Reduced lung-cancer mortality with volume CT screening in a randomized trial. N Engl J Med. 2020;382(6):503–513. doi: 10.1056/NEJMoa1911793. [DOI] [PubMed] [Google Scholar]
- 31.Henschke CI, Yankelevitz DF, Libby DM, Pasmantier MW, Smith JP, Miettinen OS, et al. Survival of patients with stage I lung cancer detected on CT screening. N Engl J Med. 2006;355(17):1763–1771. doi: 10.1056/NEJMoa060476. [DOI] [PubMed] [Google Scholar]
- 32.Hoffman RM, Atallah RP, Struble RD, Badgett RG. Lung cancer screening with low-dose CT: a meta-analysis. J Gen Intern Med. 2020;35(10):3015–3025. doi: 10.1007/s11606-020-05951-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Lee CI, Tsai EB, Sigal BM, Plevritis SK, Garber AM, Rubin GD. Incidental extracardiac findings at coronary CT: clinical and economic impact. AJR Am J Roentgenol. 2010;194(6):1531–1538. doi: 10.2214/AJR.09.3587. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Hutchinson BD, Moreira AL, Ko JP. Spectrum of subsolid pulmonary nodules and overdiagnosis. Semin Roentgenol. 2017;52(3):143–155. doi: 10.1053/j.ro.2017.06.011. [DOI] [PubMed] [Google Scholar]
- 35.Liu YC, Liang CH, Wu YJ, Chen CS, Tang EK, Wu FZ. Managing persistent subsolid nodules in lung cancer: education, decision making, and impact of interval growth patterns. Diagnostics (Basel) 2023;13(16):2674. doi: 10.3390/diagnostics13162674. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Fischbach F, Knollmann F, Griesshaber V, Freund T, Akkol E, Felix R. Detection of pulmonary nodules by multislice computed tomography: improved detection rate with reduced slice thickness. Eur Radiol. 2003;13(10):2378–2383. doi: 10.1007/s00330-003-1915-7. [DOI] [PubMed] [Google Scholar]
- 37.MacMahon H, Naidich DP, Goo JM, Lee KS, Leung ANC, Mayo JR, et al. Guidelines for management of incidental pulmonary nodules detected on CT images: from the Fleischner Society 2017. Radiology. 2017;284(1):228–243. doi: 10.1148/radiol.2017161659. [DOI] [PubMed] [Google Scholar]
- 38.Christensen J, Prosper AE, Wu CC, Chung J, Lee E, Elicker B, et al. ACR lung-RADS v2022: assessment categories and management recommendations. J Am Coll Radiol. 2024;21(3):473–488. doi: 10.1016/j.jacr.2023.09.009. [DOI] [PubMed] [Google Scholar]
- 39.Alzubaidi L, Zhang J, Humaidi AJ, Al-Dujaili A, Duan Y, Al-Shamma O, et al. Review of deep learning: concepts, CNN architectures, challenges, applications, future directions. J Big Data. 2021;8(1):53. doi: 10.1186/s40537-021-00444-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.White CS. The pros and cons of searching for extracardiac findings at cardiac CT: use of a restricted field of view is acceptable. Radiology. 2011;261(2):338–341. doi: 10.1148/radiol.11111131. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Comparison of lesion and CT scan characteristics by institutions
Univariate and multivariate mixed-effects logistic regression analyses for variables affecting differences in AI-CAD detection rate across institutions