Has computer-aided diagnosis now outperformed radiologists in the accurate localization of lung nodules?—a multicenter retrospective study

Introduction

Globally, lung cancer is one of the most frequently diagnosed cancers and the leading cause of cancer-related death, with an estimated 2 million new cases and 1.76 million deaths per year (1-3). According to cancer statistics from several countries and regions worldwide, the 5-year survival rate for stage I lung cancer patients is typically approximately 70–90%, whereas the corresponding survival rate for stage IV lung cancer patients is generally less than 10% (4). Therefore, the early detection and diagnosis of lung cancer is important.

The most common computed tomography (CT) manifestation of early-stage lung cancer is an isolated pulmonary nodule. Most pulmonary nodules are benign, but the chance of malignancy increases as the diameter of the nodules increases, reaching approximately 50% when the nodule exceeds 2 cm (5). Therefore, surgery should be carried out as soon as possible when the nodule becomes larger or when conservative treatment is shown to be ineffective during follow-up (6). The precise preoperative localization of pulmonary nodules is the key to successful intraoperative resection because it can shorten the operation time, reduce the operation risk, and maximize the protection of the patient’s lung tissue (7). A major method for the early detection of lung cancer is CT screening (8-10). Radiologists often examine many chest CT images in a single day, which places a mental burden on them, affecting the accurate localization of nodules and a lack of consistent standards (11-14). To address this problem, deep learning models trained on large-scale imaging datasets have been proposed to address this problem (15-20). Some articles have reported corresponding results in clinical applications involving the use of certain software, such as deep learning-based artificial intelligence (AI) by Lan et al. and a new computer-aided diagnosis (CAD) system for early pulmonary nodule detection by Liu et al. (15,21-29). However, most existing CAD systems are primarily designed for pulmonary nodule detection or classification rather than precise three-dimensional (3D) localization, particularly for small nodules, ground-glass nodules, and lesions adjacent to complex bronchovascular structures. Moreover, conventional deep learning models often rely on single-scale feature extraction, limiting their ability to simultaneously capture global anatomical context and fine local details. To address these limitations, we developed a CAD system based on convolutional neural networks to achieve the 3D localization of pulmonary nodules. Unlike conventional CAD systems primarily designed for nodule detection or classification, the proposed framework specifically focuses on precise 3D lobar and segmental localization of pulmonary nodules to assist surgical planning and intraoperative navigation.

The aim of this study was to compare the pulmonary nodule localization accuracy of CAD and radiologist (RAD) analysis to determine whether CAD provides superior precision. We present this article in accordance with the STARD-AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0204/rc).

Methods

Study design

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was conducted in compliance with institutional policies and relevant data protection and privacy regulations. This multicenter retrospective study was approved by the Institutional Review Boards (IRBs) of Center 1 (The First Affiliated Hospital of Anhui Medical University), Center 2 (The People’s Hospital of Bozhou), and Center 3 (Anqing Municipal Hospital). The model development cohort (n=1,100; 880 for training and 220 for internal testing) was approved by the IRBs of The First Affiliated Hospital of Anhui Medical University (Gaoxin Campus and South Campus) and Hefei Chest Hospital. The requirement for informed consent was waived by the IRBs due to the retrospective design of the study and because only pre-existing imaging data were used. All patient data were fully anonymized prior to analysis, with removal of all identifiable personal information. Potentially eligible participants were retrospectively identified from the hospital picture archiving and communication system (PACS) and electronic medical records based on chest CT examination and surgical pathology databases. Participants constituted a retrospective consecutive series of patients who met the eligibility criteria at each participating center.

Patients who met the following criteria were included: (I) underwent a 1.25 mm thin-layer CT scan within 2 weeks before surgery; (II) had pulmonary nodules with diameters between 5–30 mm; (III) underwent thoracoscopic resection of pulmonary nodules; and (IV) had a definitive diagnosis on the basis of resected specimen pathology. The exclusion criteria were as follows: (I) poor-quality chest CT images and (II) incomplete clinical data. For each patient, only the largest pulmonary nodule was selected for subsequent analysis. Data for the Center 1 study were obtained from 921 patients between December 2017 and August 2023, data for the Center 2 study were obtained from 145 patients between January 2021 and February 2024, and data for the Center 3 study were obtained from 209 patients between August 2022 and August 2023, totalling 1,275 patients (Figure 1).

Figure 1 Flowchart of case collection and exclusion. Center 1, The First Affiliated Hospital of Anhui Medical University; Center 2, The People's Hospital of Bozhou; Center 3, Anqing Municipal Hospital. CT, computed tomography.

Instrumentation and inspection

The study data were obtained from a total of 1,275 chest CT image sets from three centers between December 2017 and February 2024. Center 1 scanned patients using uCT 760, uCT 780 (United Imaging, Shanghai, China), and Revolution CT (GE HealthCare, Waukesha, WI, USA) scanners; Center 2 scanned patients using SOMATOM Definition Flash (Siemens Medical Systems, Forchheim, Germany) and Revolution CT scanners; and Center 3 scanned patients using SOMATOM Definition Flash, uCT 530 (United Imaging), and NeuViz 64 CT (Neusoft, Shenyang, China) scanners (Table 1). All study cases were reconstructed with a slice thickness of 1.25 mm, which is a commonly used thin-section CT protocol for pulmonary nodule evaluation and provides sufficient spatial resolution for pulmonary nodule localization and pulmonary segment analysis. All CT images were acquired using standard clinical scanning protocols. No uniform contrast enhancement protocol was applied, as both contrast-enhanced and non-contrast scans were included according to routine clinical practice. No additional image preprocessing (e.g., resampling or intensity normalization) was applied prior to input into the CAD system.

Group setting and classification of nodal characteristics

A total of 1,275 chest CT images from three centers were analysed using the proposed CAD system to obtain the lobar and segmental distributions of the pulmonary nodules. For the RAD assessment, three professional and experienced radiologists with more than 5 years of diagnostic experience in chest imaging independently reviewed all CT images to identify the lobar and segmental distributions of the pulmonary nodules. Pulmonary nodules were manually localized according to segmental lung anatomy using axial CT images in combination with multiplanar reconstruction (MPR). To eliminate reading order bias, the cases were presented in a predetermined randomized order, and each reader read them in that order. No clinical, demographic, or database information about the patients was provided to the readers during the study. The gold standard for nodule localization was established based on surgical findings, postoperative histopathological confirmation of the resected specimens, and correlation with preoperative CT images. The lobar and segmental locations of the nodules were determined by surgeons and radiologists according to preoperative CT anatomical landmarks and operative records. Surgeons pinpointed nodules to the pulmonary segment with sufficient precision for this study. All relevant staff followed unified pulmonary segment classification criteria for reporting and completed standardized pre-study training to guarantee consistent descriptions.

CAD system

The CAD system is based on a 3D U-Net deep learning model. A simple workflow scheme for model development is shown in Figure 2. The CAD system was developed as a research software tool; it uses 3D U-Net deep learning models to detect and segment pulmonary nodules for subsequent lobar and segmental localization. It was designed to assist thoracic radiologists and thoracic surgeons in pulmonary nodule localization and pulmonary segment identification. The algorithmic pipeline comprises three consecutive modules: (I) lung parenchyma, lobar, and nodular segmentation; (II) bronchovascular tree segmentation; and (III) bronchial anomaly detection with pulmonary segment partitioning. The first 3D U-Net architecture was implemented for simultaneous segmentation of the pulmonary parenchyma, lung lobes, and nodules. Isolating the lung volume through parenchymal segmentation not only eliminates interference from extrapulmonary organs but also optimizes computational efficiency. Subsequent lobar segmentation delineates interlobar fissures to demarcate boundaries between the ipsilateral and contralateral lobes, whereas nodular segmentation precisely localizes lesion margins. This hierarchical approach establishes spatial priors for subsequent segmental analysis. The second 3D U-Net was dedicated to bronchovascular segmentation, with particular emphasis on preserving topological fidelity. Given that intersegmental planes are anatomically defined by bronchial bifurcations and require intact vascular supply within segments, this module provides the morphological foundation for segmental division by reconstructing bronchial generations (up to the 4th–5th order) and their accompanying vascular networks. The third 3D U-Net addresses anatomical variations through bifurcation-aware anomaly detection. To mitigate segmentation errors caused by aberrant bronchial branching beyond standard segmental territories, we introduced Gaussian spheres to model bifurcation probability heatmaps. These spatially encoded priors, concatenated with raw CT inputs, enabled the robust detection of variant segments. Integration of all the hierarchical features ultimately achieved high-precision pulmonary segment partitioning (Figure 3). During the development and training of the product model, single-scan imaging data from 1,100 distinct patients across three additional medical centers were partitioned into training (n=880) and test sets (n=220) at an 8:2 ratio (Table S1). The training dataset and validation dataset were completely independent, with no patient overlap. Although the test cohort was derived from the same multicenter dataset rather than an independent external cohort, the dataset included cases from three different medical centers, multiple CT scanner vendors (Siemens, GE HealthCare, Philips, and Toshiba/Canon), and heterogeneous acquisition protocols. This diversity enabled evaluation of the model under varying clinical and imaging conditions and provided an assessment of its robustness across different clinical settings. In the multicenter validation involving 220 cases, the model demonstrated robust segmentation performance, with Dice similarity coefficients of 0.970 for bronchial structures, 0.954 for pulmonary arteries/veins, 0.958 for pulmonary segments, and 0.989 for pulmonary nodules (Table S2).

Figure 2 Schematic diagram of the CAD system. 3D, three-dimensional; CAD, computer-aided diagnosis; conv, convolutional layer; CT, computed tomography; IN, instance normalization; IReLU, Inplace Rectified Linear Unit.

Figure 3 A 58-year-old man with a 12-mm lung adenocarcinoma. (A,B) Computed tomography scan showing a tumour located in the left S5. (C,D) Digital 3D model with pulmonary artery (blue), pulmonary vein (pink), bronchial (white), tangent sphere (yellow), and pulmonary nodules (yellow-green) in S5, S1+2 (light tan), S3 (off-white), S4, S5 (salmon pink), S6 (light purple), S7+8 (gray-green), S9 (green), S10 (slate gray), which shows the relationships between pulmonary nodules and the bronchial and arteriovenous systems.

Statistical analysis

Categorical variables are expressed as absolute numbers and percentages, and comparisons between groups were performed using the chi-square test or Fisher’s exact test. All the statistical tests were two-sided, with P<0.05 indicating a statistically significant difference. Data analysis was performed using SPSS software (version 27.0). Dice similarity coefficients were calculated on a per-patient basis and reported as mean values with 95% confidence intervals (CIs).

Results

Demographic and imaging characteristics

The demographic data and imaging characteristics of patients from the three different medical centers are summarized in Table 2. A total of 1,275 patients were recruited from three centers, including 921 patients in Center 1, 145 in Center 2, and 209 in Center 3. A total of 15,311 pulmonary nodules were detected among these enrolled patients, with 12,055, 1,777, and 1,479 nodules from the three centers, respectively. We compared the clinical characteristics of patients from the three centers in terms of age, sex, nodule size, nodule type, nodule characteristics, surgical approach, pathological findings, and nodule location, and there were no statistically significant differences in the clinical characteristics among patients from the three centers (all P>0.05).

Two-by-two comparison of the gold standard, CAD and RAD

In these three centers, the accuracy rates of pulmonary nodule localization for CAD versus the gold standard, RAD versus the gold standard, and CAD versus RAD across lobes and segments of the lungs were compared.

In Center 1 (n=921), CAD correctly located 856 pulmonary nodules, with an accuracy rate of 92.9% (95% CI, 91.1–94.4%), which exceeded 90% for each lobe and reached 100% for the right middle lobe of the lung. RAD correctly located 758 nodules, with accuracy rates of 82.3% (95% CI, 79.7–84.6%) (P<0.001). Notably, the accuracy rate did not exceed 90% for any lobe. The overall accuracy rate for CAD was better than that for RAD (P<0.001). In addition, except for the middle lobe of the right lung, the accuracy rates of CAD were better than those of RAD in the remaining lobes of the lung (all P<0.05) (Table 3).

In Center 2 (n=145), CAD correctly located 141 pulmonary nodules, with an accuracy rate of 97.2% (95% CI, 94.54–99.90%). The accuracy rate of each lung lobe was more than 90%, with the accuracy rate of the right upper lobe and the right middle lobe of the lung reaching 100%. A total of 112 nodules were correctly located by RAD, with an accuracy rate of 77.2% (95% CI, 70.39–84.05%) (P<0.001). The accuracy rates of the right upper lobe and right lung middle lobe were greater than 80%, and those of each of the remaining lobes were less than 80%. The overall accuracy rate for CAD was better than that for RAD (P<0.001). In addition, the accuracy rates for CAD were better than those for RAD in the remaining lobes of the lungs, except for the right middle lobe and the right lower lobe (all P<0.05) (Table 4).

In Center 3 (n=209), CAD correctly located 196 pulmonary nodules, with an accuracy detection rate of 93.8% (95% CI, 90.53–97.07%). The accuracy rate for each lung lobe exceeded 90%, and the accuracy rate for the right middle lobe of the lung reached 100%. RAD correctly located 163 nodules, with an accuracy rate of 78% (95% CI, 72.38–83.62%) (P<0.001). The accuracy rate of the left upper lobe of the lung was greater than 80%, and that of each of the remaining lobes was less than 80%. The overall accuracy rate for CAD was better than that for RAD (P<0.001). In addition, except for the middle lobe of the right lung, the accuracy rates of CAD in each lobe of the lung were better than those of RAD (all P<0.05) (Table 5).

With CAD, the average accuracy rates for pulmonary nodule localization of different lung segments improved across all 3 centers (P<0.001) (Figure 4).

Figure 4 The average accuracy of different lung segments in the 3 centers under two reading conditions: without CAD and with CAD. CAD, computer-aided diagnosis.

Discussion

With the extensive application of CT in clinical diagnostics, the workload of radiologists has increased dramatically, resulting in increased fatigue and a phenomenon known as “satisfaction of search”, which refers to the fact that in a CT image of the same individual, the presence of a prominent lesion can attract the visual attention of a radiologist, potentially leading to overlooking other pulmonary nodules; consequently, the physician may conclude the review of the images before all the nodules are detected. These factors substantially diminish the precision of pulmonary nodule diagnosis (27,28). To address this issue, we have implemented a CAD system as an auxiliary diagnostic tool. This system automatically detects, localizes, segments, and quantifies pulmonary nodules, thereby reducing workload and enhancing diagnostic consistency. Recent studies have demonstrated that CAD systems can support all stages of pulmonary nodule management, including detection, malignancy risk assessment, diagnosis, and follow-up monitoring (26-30). Several studies have explored a range of related predictive tasks aimed at more comprehensive lesion characterization, among which Usman et al. (17-19) developed a series of multi-encoder and attention-based models. They first proposed an adaptive region of interest (ROI)-based volumetric segmentation method combined with multi-view residual learning to improve segmentation performance by dynamically optimizing regions of interest. They then designed a multi-encoder self-distilled detection network integrated with bidirectional maximum intensity projections to enhance detection sensitivity and reduce false positives. Finally, they advanced a multi-encoder network with a self-adaptive hard attention mechanism to strengthen small nodule feature representation and eliminate rescaling-induced errors (17-19). Therefore, CAD has the potential to improve the accuracy of pulmonary nodule localization, and its integration into daily clinical practice is important. In addition, CAD is more consistent than radiologists, thereby avoiding interobserver variability, as well as more efficient and able to save more labour (21,25,29).

While existing CAD systems improve nodule detection consistency and reduce interobserver variability (30,31), their capacity to enable precision thoracoscopic surgery remains limited by two critical gaps: (I) inadequate subsegmental localization accuracy owing to anatomical variations (32,33); (II) compromised generalizability across multicenter datasets (22). Although the primary objective of the CAD system was pulmonary nodule localization, lobar and vascular segmentation were incorporated to provide anatomical context for model learning. Because pulmonary nodules are often located adjacent to vessels or fissures, segmentation of surrounding anatomical structures may improve differentiation between nodules and adjacent tissues, reduce false-positive findings, and enhance localization accuracy. In addition, automated visualization of lobar and bronchovascular anatomy may provide useful information for preoperative assessment and surgical planning. Existing CAD systems mainly focus on pulmonary nodule detection and malignancy prediction, whereas accurate 3D localization remains challenging, particularly for small nodules and lesions adjacent to complex bronchovascular structures. Precise localization requires not only nodule segmentation but also accurate characterization of surrounding anatomical structures, including pulmonary vessels and segmental anatomy. Our three-tier 3D U-Net architecture fundamentally addresses these limitations through hierarchical task decomposition: (I) lobar-level segmentation with interlobar fissure delineation eliminates cardiac diaphragm interference, achieving 100% sensitivity in the right middle lobe, which outperforms the 76–85% range reported in the literature (34); (II) Bifurcation-aware anomaly detection using Gaussian probability heatmaps reduces bronchovascular segmentation errors by 37% compared with conventional nnU-Net (Dice: 0.958 vs. 0.89; P<0.001) (22); (III) multicenter adaptive normalization [intensity/size alignment + ANTs (Advanced Normalization Tools)/CLAHE (contrast-limited adaptive histogram equalization)] maintains >93% sensitivity across heterogeneous scanners, in contrast with >15% performance drops in single-center models (26).

Compared with previous studies, our study showed greater accuracy in the localization of pulmonary nodules. For example, in the study by Kim et al., the mean area under the curve of readers increased from 0.82 to 0.89 when CAD was used (27). In the work of Chao et al., AI was able to help radiologists, and the overall sensitivity of detecting different nodules increased from 67.7% to 88.4% (33).

In our study, the accuracy rate of pulmonary nodule localization in centers 1, 2, and 3 increased from 82.3%, 77.2%, and 78.0% for RAD to 92.9%, 97.2%, and 93.8% for CAD, respectively. On the basis of these results, we hypothesized that the CAD system we developed could assist in the 3D localization of pulmonary nodules, which may help clinicians accurately locate nodules during the operation, reduce the operation time, and minimize the damage to adjacent tissues. However, the difference in accuracy between CAD and RAD in the right middle lobe of the lungs in all three imaging centers was not significant, possibly because of the lower incidence of nodal lesions in the right middle lobe of the lungs in the clinic. In a study by Tietz (35), the incidence of nodules in the right middle lobe was lower than that in the other lobes of the lungs, possibly because of the smaller size of the right middle lobe, which naturally leads to a smaller number of nodules. Alternatively, it was difficult to detect nodules in the middle lobe of the right lung because of occlusion of the cardiac shadow or low contrast of neighbouring structures. Therefore, more samples of right middle lobe lesions need to be included in future studies to increase performance in this lung segment (34).

Our study has several limitations. First, our CAD system failed to provide the reader with essential clinical background information regarding the patients, which limits the clinical generalizability of the results to a certain extent. However, this design was chosen to avoid bias resulting from the clinical background and to ensure that the analysis results are mainly based on image features. Second, in this study, pulmonary nodules were divided into two groups (5–<10 mm and 10–30 mm) according to the predefined research scheme. Since the primary objective was to evaluate the overall performance of the CAD system for nodules within the 5–30 mm range, we did not further subdivide the size categories into three subgroups. Consistent with previous clinical findings and our observational results, the detection and localization performance of the CAD system was generally better for larger pulmonary nodules. Third, one limitation of this study is that the proposed CAD system was not evaluated on public benchmark datasets such as LIDC-IDRI. The LIDC-IDRI dataset mainly provides annotations for pulmonary nodule detection and benign-malignant discrimination, which does not match our research focus on pulmonary segment and lobe localization. Future work will introduce external public datasets to conduct benchmark tests, so as to realize intuitive comparative analysis with existing relevant studies. Fourth, although the model was developed and evaluated using a multicenter dataset with heterogeneous imaging protocols and scanner vendors, an independent external validation cohort was not available. Therefore, the generalizability of the proposed CAD system should be further confirmed in future studies using completely independent datasets from additional institutions. Fifth, as with all CAD-based studies, the results of this study are applicable only to the specific CAD system used and should not be directly generalized to other systems. Therefore, in future studies, side-by-side comparisons between different systems need to be performed to clarify the applicability and limitations of different algorithms. Future studies could consider applying the CAD system to larger and diverse clinical datasets, especially in combination with clinical contexts in real diagnostic environments, to further enhance its generalizability. This study is expected to further improve the application value and diagnostic efficiency of CAD systems in the clinic and to lay a foundation for the further development of precision medicine.

Conclusions

In summary, this study demonstrated that the proposed AI-based CAD system provided a higher localization accuracy rate in the diagnosis of pulmonary nodules than manual detection by radiologists.

Acknowledgments

This multicenter retrospective study relied on the generous sharing of clinical resources. We extend our appreciation to the teams at The People’s Hospital of Bozhou and Anqing Municipal Hospital for their pivotal role in case collection.

Footnote

Reporting Checklist: The authors have completed the STARD-AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0204/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0204/dss

Funding: This study was supported by Guangzhou National Laboratory (No. GZNL2023A03001).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0204/coif). L.L., D.G., Z.Y., and Y.H. are employees of Shukun (Beijing) Network Technology Co., Ltd. J.L. is an employee of GE HealthCare China. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was conducted in compliance with institutional policies and relevant data protection and privacy regulations. This multicenter retrospective study was approved by the Institutional Review Boards (IRBs) of The First Affiliated Hospital of Anhui Medical University, The People’s Hospital of Bozhou, and Anqing Municipal Hospital. The model development cohort (n=1,100) was approved by the IRBs of The First Affiliated Hospital of Anhui Medical University (Gaoxin Campus and South Campus) and Hefei Chest Hospital. The requirement for informed consent was waived by the IRBs due to the retrospective design of the study and because only pre-existing imaging data were used. All patient data were fully anonymized prior to analysis, with removal of all identifiable personal information.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Brody H. Lung cancer. Nature 2020;587:S7. [Crossref] [PubMed]
2. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
3. Thai AA, Solomon BJ, Sequist LV, Gainor JF, Heist RS. Lung cancer. Lancet 2021;398:535-54. [Crossref] [PubMed]
4. Maghfoor I, Perry MC. Lung cancer. Ann Saudi Med. 2005;25:1-12. [Crossref] [PubMed]
5. Ocaña-Tienda B, Eroles-Simó A, Pérez-Beteta J, Arana E, Pérez-García VM. Growth dynamics of lung nodules: implications for classification in lung cancer screening. Cancer Imaging 2024;24:113. [Crossref] [PubMed]
6. Dziedzic R, Marjański T, Rzyman W. A narrative review of invasive diagnostics and treatment of early lung cancer. Transl Lung Cancer Res 2021;10:1110-23. [Crossref] [PubMed]
7. Nardini M, Dunning J. Pulmonary nodules precision localization techniques. Future Oncol 2020;16:15-9. [Crossref] [PubMed]
8. Lv X, Lian Y, Liu Z, Xiao J, Zhang D, Yin X. Exosomal long non-coding RNA LINC00662 promotes non-small cell lung cancer progression by miR-320d/E2F1 axis. Aging (Albany NY) 2021;13:6010-24. [Crossref] [PubMed]
9. Mazzone PJ, Lam L. Evaluating the Patient With a Pulmonary Nodule: A Review. JAMA 2022;327:264-73. [Crossref] [PubMed]
10. Aberle DR, Adams AM, Berg CD, Black WC, Clapp JD, Fagerstrom RM, Gareen IF, Gatsonis C, Marcus PM, Sicks JD. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365:395-409.
11. Peloschek P, Sailer J, Weber M, Herold CJ, Prokop M, Schaefer-Prokop C. Pulmonary nodules: sensitivity of maximum intensity projection versus that of volume rendering of 3D multidetector CT data. Radiology 2007;243:561-9. [Crossref] [PubMed]
12. Field JK, Oudkerk M, Pedersen JH, Duffy SW. Prospects for population screening and diagnosis of lung cancer. Lancet 2013;382:732-41. [Crossref] [PubMed]
13. Wiener RS, Gould MK, Slatore CG, Fincke BG, Schwartz LM, Woloshin S. Resource use and guideline concordance in evaluation of pulmonary nodules for cancer: too much and too little care. JAMA Intern Med 2014;174:871-80. [Crossref] [PubMed]
14. Zhang C, Sun X, Dang K, Li K, Guo XW, Chang J, et al. Toward an Expert Level of Lung Cancer Detection and Classification Using a Deep Convolutional Neural Network. Oncologist 2019;24:1159-65. [Crossref] [PubMed]
15. Liu JK, Jiang HY, Gao MD, He CG, Wang Y, Wang P, Ma H, Li Y. An Assisted Diagnosis System for Detection of Early Pulmonary Nodule in Computed Tomography Images. J Med Syst 2017;41:30. [Crossref] [PubMed]
16. Jeon KN, Goo JM, Lee CH, Lee Y, Choo JY, Lee NK, Shim MS, Lee IS, Kim KG, Gierada DS, Bae KT. Computer-aided nodule detection and volumetry to reduce variability between radiologists in the interpretation of lung nodules at low-dose screening computed tomography. Invest Radiol 2012;47:457-61. [Crossref] [PubMed]
17. Usman M, Lee BD, Byon SS, Kim SH, Lee BI, Shin YG. Volumetric lung nodule segmentation using adaptive ROI with multi-view residual learning. Sci Rep 2020;10:12839. [Crossref] [PubMed]
18. Usman M, Rehman A, Shahid A, Latif S, Byon SS, Lee BD, Kim SH, Lee BI, Shin YG. MEDS-Net: Self-Distilled Multi-Encoders Network with Bi-Direction Maximum Intensity projections for Lung Nodule Detection. arXiv [Preprint]. 2022. arXiv: 2211.00003.
19. Usman M, Rehman A, Ur Rehman A, Shahid A, Khan TM, Razzak I, Chung M, Shin YG. Multi-encoder self-adaptive hard attention network with maximum intensity projections for lung nodule segmentation. Comput Biol Med 2025;197:111059. [Crossref] [PubMed]
20. Ahmad IS, Dai J, Xie Y, Liang X. Deep learning models for CT image classification: a comprehensive literature review. Quant Imaging Med Surg 2025;15:962-1011. [Crossref] [PubMed]
21. Lv J, Li J, Liu Y, Zhang H, Luo X, Ren M, Gao Y, Ma Y, Liang S, Yang Y, Song Z, Gao G, Gao G, Jiang Y, Li X. Artificial Intelligence-Aided Diagnosis Software to Identify Highly Suspicious Pulmonary Nodules. Front Oncol 2021;11:749219. [Crossref] [PubMed]
22. Murchison JT, Ritchie G, Senyszak D, Nijwening JH, van Veenendaal G, Wakkie J, van Beek EJR. 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:e0266799. [Crossref] [PubMed]
23. Kim RY, Oke JL, Pickup LC, Munden RF, Dotson TL, Bellinger CR, Cohen A, Simoff MJ, Massion PP, Filippini C, Gleeson FV, Vachani A. Artificial Intelligence Tool for Assessment of Indeterminate Pulmonary Nodules Detected with CT. Radiology 2022;304:683-91. [Crossref] [PubMed]
24. Smith D, Melville P, Fozzard N, Zhang J, Deonarine P, Nirthanan S, Sivakumaran P. Artificial intelligence software in pulmonary nodule assessment. J R Coll Physicians Edinb 2022;52:228-31. [Crossref] [PubMed]
25. Ni YL, Zheng XC, Shi XJ, Xu YF, Li H. Deep convolutional neural network based on CT images of pulmonary nodules in the lungs of adolescent and young adult patients with osteosarcoma. Oncol Lett 2023;26:344. [Crossref] [PubMed]
26. Jiang L, Zhou Y, Miao W, Zhu H, Zou N, Tian Y, Pan H, Jin W, Huang J, Luo Q. Artificial intelligence-assisted quantitative CT parameters in predicting the degree of risk of solitary pulmonary nodules. Ann Med 2024;56:2405075. [Crossref] [PubMed]
27. Kim RY, Oke JL, Dotson TL, Bellinger CR, Vachani A. Effect of an artificial intelligence tool on management decisions for indeterminate pulmonary nodules. Respirology 2023;28:582-4. [Crossref] [PubMed]
28. Zhang L, Shao Y, Chen G, Tian S, Zhang Q, Wu J, Bai C, Yang D. An artificial intelligence-assisted diagnostic system for the prediction of benignity and malignancy of pulmonary nodules and its practical value for patients with different clinical characteristics. Front Med (Lausanne) 2023;10:1286433. [Crossref] [PubMed]
29. Lan CC, Hsieh MS, Hsiao JK, Wu CW, Yang HH, Chen Y, Hsieh PC, Tzeng IS, Wu YK. Deep Learning-based Artificial Intelligence Improves Accuracy of Error-prone Lung Nodules. Int J Med Sci 2022;19:490-8. [Crossref] [PubMed]
30. Berbaum KS, Franken EA Jr, Dorfman DD, Rooholamini SA, Kathol MH, Barloon TJ, Behlke FM, Sato Y, Lu CH, el-Khoury GY. Satisfaction of search in diagnostic radiology. Invest Radiol 1990;25:133-40. [Crossref] [PubMed]
31. Jin H, Yu C, Zhang J, Zheng R, Fu Y, Zhao Y. Multitask Swin Transformer for classification and characterization of pulmonary nodules in CT images. Quant Imaging Med Surg 2025;15:1845-61. [Crossref] [PubMed]
32. Cao K, Meng G, Wang Z, Liu Y, Liu H, Sun L. An adaptive pulmonary nodule detection algorithm. J Xray Sci Technol 2020;28:427-47. [Crossref] [PubMed]
33. Chao HS, Tsai CY, Chou CW, Shiao TH, Huang HC, Chen KC, Tsai HH, Lin CY, Chen YM. Artificial Intelligence Assisted Computational Tomographic Detection of Lung Nodules for Prognostic Cancer Examination: A Large-Scale Clinical Trial. Biomedicines 2023;11:147. [Crossref] [PubMed]
34. Araujo-Filho JAB, Halpenny D, McQuade C, Puthoff G, Chiles C, Nishino M, Ginsberg MS. Management of Pulmonary Nodules in Oncologic Patients: AJR Expert Panel Narrative Review. AJR Am J Roentgenol 2021;216:1423-31. [Crossref] [PubMed]
35. Tietz E, Müller-Franzes G, Zimmermann M, Kuhl CK, Keil S, Nebelung S, Truhn D. Evaluation of Pulmonary Nodules by Radiologists vs. Radiomics in Stand-Alone and Complementary CT and MRI. Diagnostics (Basel) 2024;14:483.