Accuracy of AI-Based Algorithms in Pulmonary Embolism Detection on Computed Tomographic Pulmonary Angiography: An Updated Systematic Review and Meta-analysis.

Accuracy of AI-Based Algorithms in Pulmonary Embolism Detection on Computed Tomographic Pulmonary Angiography: An Updated Systematic Review and Meta-analysis - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Imaging Inform Med . 2025 Sep 15;39(3):2239–2253. doi: 10.1007/s10278-025-01645-w Accuracy of AI-Based Algorithms in Pulmonary Embolism Detection on Computed Tomographic Pulmonary Angiography: An Updated Systematic Review and Meta-analysis Seyed Ali Nabipoorashrafi 1 , Arsalan Seyedi 2 , Razman Arabzadeh Bahri 3 , Amirhossein Yadegar 2 , Mostafa Shomal-Zadeh 2 , Fatemeh Mohammadi 2 , Samira Amin Afshari 2 , Negar Firoozeh 4 , Navida Noroozzadeh 5 , Farbod Khosravi 1 , Sanaz Asadian 1 , Hamid Chalian 1, ✉ Author information Article notes Copyright and License information 1 Cardiothoracic Imaging Section, Department of Radiology, University of Washington, 1959 NE Pacific Street Room RR215F, Box 357115, Seattle, WA 98195 USA 2 Endocrinology and Metabolism Research Center (EMRC), School of Medicine, Vali-Asr Hospital, P.O. Box 13145784, Tehran, Iran 3 Urology Research Center, Tehran University of Medical Sciences, Tehran, Iran 4 Department of Radiology, University of Washington School of Medicine, Seattle, WA USA 5 School of Medicine, Alborz University of Medical Sciences, Karaj, Alborz Iran ✉ Corresponding author. Received 2025 May 6; Revised 2025 Aug 6; Accepted 2025 Aug 7; Collection date 2026 Jun. © The Author(s) under exclusive licence to Society for Imaging Informatics in Medicine 2025. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC13230453  PMID: 40954364 Abstract Several artificial intelligence (AI) algorithms have been designed for detection of pulmonary embolism (PE) using computed tomographic pulmonary angiography (CTPA). Due to the rapid development of this field and the lack of an updated meta-analysis, we aimed to systematically review the available literature about the accuracy of AI-based algorithms to diagnose PE via CTPA. We searched EMBASE, PubMed, Web of Science, and Cochrane for studies assessing the accuracy of AI-based algorithms. Studies that reported sensitivity and specificity were included. The R software was used for univariate meta-analysis and drawing summary receiver operating characteristic (sROC) curves based on bivariate analysis. To explore the source of heterogeneity, sub-group analysis was performed (PROSPERO: CRD42024543107). A total of 1722 articles were found, and after removing duplicated records, 1185 were screened. Twenty studies with 26 AI models/population met inclusion criteria, encompassing 11,950 participants. Univariate meta-analysis showed a pooled sensitivity of 91.5% (95% CI 85.5–95.2) and specificity of 84.3 (95% CI 74.9–90.6) for PE detection. Additionally, in the bivariate sROC, the pooled area under the curved (AUC) was 0.923 out of 1, indicating a very high accuracy of AI algorithms in the detection of PE. Also, subgroup meta-analysis showed geographical area as a potential source of heterogeneity where the I 2 for sensitivity and specificity in the Asian article subgroup were 60% and 6.9%, respectively. Findings highlight the promising role of AI in accurately diagnosing PE while also emphasizing the need for further research to address regional variations and improve generalizability. Supplementary Information The online version contains supplementary material available at 10.1007/s10278-025-01645-w. Keywords: Pulmonary embolism, Artificial intelligence, Tomography, X-ray computed Introduction Pulmonary embolism (PE) remains a significant cause of morbidity and mortality worldwide, with an estimated incidence of 60 to 70 cases per 100,000 population annually [ 1 ]. Additionally, sub-segmental PE yields a lower mortality rate of 20 cases per 10,000; however, significant risk of recurrence and major bleeding remain [ 2 ]. Early and accurate diagnosis of PE is crucial for guiding appropriate treatment decisions and improving patient outcomes. Computed tomographic pulmonary angiography (CTPA) is widely regarded as the imaging modality of choice for diagnosing PE due to its noninvasiveness, high sensitivity, and high specificity [ 3 ]. This non-invasive procedure can quickly and accurately identify the presence, location, and extent of PE, helping to guide appropriate treatment decisions and improve patient outcomes [ 3 , 4 ]. Historically, the development of AI in PE detection has been marked by a shift from rule-based and feature-engineered approaches to data-driven deep learning models, which can automatically extract complex imaging features and patterns associated with PE [ 5 , 6 ]. Artificial intelligence (AI) algorithms have shown the ability to analyze complex imaging data rapidly and precisely, potentially lowering interpretation time and enhancing sensitivity and specificity of PE detection [ 7 ]. Recent studies have investigated various AI-based approaches for PE diagnosis using CTPA, including machine learning algorithms, and deep learning neural networks systems [ 8 ]. These approaches leverage advanced image processing techniques to extract quantitative features from CTPA images and identify patterns indicative of PE, thereby enhancing the radiologists’ ability in PE detection, by increasing sensitivity and specificity, reducing missed diagnosis, prioritizing cases and enhancing workflow, reducing diagnostic errors, decision support, and faster interpretation. However, recent studies on adversarial robustness have revealed the concerning fragility of AI systems to noise and disturbances in CT imaging, which poses considerable safety concerns for PE detection, which could lead to misclassification or oversight of clinically important findings. Consequently, there is an urgent need for strong defenses and thorough validation of AI models to guarantee reliable and secure implementation in clinical CT imaging processes for PE identification [ 9 , 10 ]. Composing AI algorithms and methods, which merge multiple neural network models, improves the accuracy. These methods aggregate predictions from different networks, lowering the likelihood of false negatives and false positives. In general, AI integration in CTPA for PE detection warrants major progress in the field of radiology, providing accurate, consistent, and prompt diagnostic support. The most common AI algorithms being used for PE detection in imaging are convolutional neural networks (CNNs) and their variants, including CNN-UNet, CNN-PENet, and deep learning-CNN [ 11 ]. In the present report, we aimed to provide a systematic review and meta-analysis of the current AI-based approaches for PE detection using CTPA and investigate recent studies in AI-based CTPA analysis for PE detection, emphasizing the benefits and limitations of these approaches as well as the sensitivity and specificity of AI algorithms for detection of PE. Methods The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement 2020 guideline was used for conducting this systematic review. The method of this study was registered in the International Prospective Register of Systematic Reviews (PROSPERO) with the identifier CRD42024543107. Search Strategy Four online databases, including Embase, Web of Science, PubMed, and Cochrane, were systematically searched until June 2023 using terms related to “artificial intelligence” and “pulmonary embolism” (Supplementary file). Study Selection In this study, all original peer-reviewed articles that evaluated AI algorithms for detecting PE in human subjects using any kind of imaging modality were included. Sensitivity, specificity, and sample size models were required as inclusion criteria. However, if a study reported data from which sensitivity, specificity, and sample size could be calculated, it was also included (e.g., true positive (TP), true negative (TN), false positive (FP), and false negative (FN)). Also, if the study used computer-aided detection (CAD) algorithms and did not specify whether they were AI-based or not, the article was included. We excluded articles that were preprints, review articles, projects on incidental PE, and those with algorithms that were clearly not AI-based. Data Extraction A predesigned Google Sheet was used to extract data from the included articles. The extracted data included the first author’s name, year of publication, location where the research was conducted (country), study design, publication journal, mean age of participants, name of the AI model used (or only CAD if it did not report whether the CAD is AI-based or not), sample size (number of cases and controls), TN, TP, FP, FN, sensitivity, specificity, female to male ratio, and the slice thickness of CT for detection by AI. Data for each included article were extracted by two independent contributors, with any disagreements resolved through discussion between the two extractors; if unresolved, a third contributor extracted the specific data from the main article. Analysis The meta-analysis was conducted using two different approaches: Univariate approach: The pooled values of various diagnostic test accuracy (DTA), such as sensitivity and specificity, were calculated separately (a random effects model was utilized for this analysis). Bivariate approach: Sensitivity and specificity were jointly pooled, representing a more recent method compared to the univariate approach. The area under the summary receiver operating characteristic (sROC) curve (AUC), which ranges from zero to one, provides an overall assessment of the diagnostic accuracy of AI algorithms, with higher values indicating better performance. To evaluate the heterogeneity between studies in the meta-analyses of DTA, the Cochran- Q test, τ 2 , and I 2 index were employed. Significant heterogeneity of analysis was defined as I 2 > 50% or a significant Cochran- Q test ( P < 0.10). Analyses were performed as follows: (1) on all included studies, (2) on studies explicitly mentioning AI algorithms (CAD algorithms that were not explicitly stated to use AI were removed), and (3) on studies explicitly mentioning AI algorithms on external populations (CAD algorithms that were not explicitly stated to use AI were removed, and AI algorithms validated on internal populations were also removed). To evaluate the effects of possible confounding factors on the heterogeneity among publications, subgroup analyses were performed based on the geographical area of the article, year of publication, and slice thickness of CT scan. Additionally, based on the AI algorithm architecture, we performed a bivariate meta-analysis to evaluate heterogeneity and differences in results according to the type of AI used. For this purpose, we categorized all reported AI architectures into one of the following main groups: classic and advanced CAD, pure DL, and specialized CNN. Quality Assessment For the evaluation of study quality, the QUADAS-2 tool was utilized. This tool assesses studies across four different domains: patient selection, index test, reference standard, and flow and timing. Each domain consists of three signaling questions, scored as yes/no/unclear, indicating low risk of bias, high risk of bias, or inability to assess bias due to insufficient information. A “yes” response to all signaling questions within a domain signifies low bias risk, while a “no” response suggests high bias risk, and “unclear” indicates insufficient data. Additionally, QUADAS-2 addresses applicability concerns through three questions. Any disagreements were resolved through discussion with other reviewers. We used the graphical display template from the University of Bristol to visually represent the QUADAS-2 assessment. Also, publication bias was evaluated with funnel plot, Egger’s, and Begg’s. Result Study Characteristics After removing duplicate records, 1185 articles were screened by two independent reviewers by title and abstract. At this step, the articles were screened conservatively (801 articles were removed based on title/abstract screening). In the next step, the full manuscripts of the articles included from the previous step were reviewed again by two independent reviewers (325 articles were removed based on full text screening). Any disagreements regarding the inclusion or exclusion of articles were resolved by discussion between reviewers or by a third reviewer. After excluding 38 articles due to lack of sufficient data for meta-analysis, 21 articles were finally included [ 8 , 12 , 13 ] (Fig. 1 ). Fig. 1. Open in a new tab PRISMA chart Considering that five articles provided accuracy data on more than one population and/or algorithms, 26 different data sets were included in the meta-analysis [ 14 – 18 ]. A total of 11,950 individuals, including 4030 patients with PE and 7920 without PE, were included in this analysis. The sample size of the included studies ranged from 36 to 1770 (Table 1 ). The highest reported sensitivity was 100% in four articles (Schoepf, 2007, E; Liu, 2019, E; Ebrahimian [sub-optimal quality]; and Zhang) [ 15 , 18 – 20 ], while the highest specificity was for Jacob, 2022, E [ 12 ] with 99.3%, also, the only modality that was used for detection of PE in studies was the CT scan (Table 1 ). Table 1. Characteristics of included studies First author Year Validation Study design Algorithm Country Image modality Total number of participants Case number (diagnosed) Control number Sensitivity Specificity Slice thickness (mm) Baxa 2008 External Retrospective CAD Czech Republic CTs 36 18 18 83.3 61 NR Blackmon 2011 External Retrospective CAD USA CTs 79 36 43 93.75 14.8 NR Huang 2020 Internal Retrospective PENet USA CTs 163 85 78 73.0 82.0 1.25 Huang 2020 External Retrospective PENet USA CTs 198 94 104 75.0 80.0 2 Huang 2020 Internal Retrospective PENet USA CTs 190 110 80 55.9 86.2 1.25 Jacob 2022 External Retrospective DCNN (Aidoc) Netherlands CTs 1275 100 1175 99.0 99.3 NR Muller-Peltzer 2021 External Retrospective CAD Deutschland CTs 1229 182 1047 64.8 21.6 0.75 Walsham 2007 External Retrospective CAD Canada CTs 100 21 79 86.0 20 1 Schoepf 2007 External Retrospective CAD USA CTs 36 23 13 100 89.9 1 Grenier 2023 External Retrospective Unet USA CTs 387 186 201 91.4 91.5 ≤ 2.5 Grenier 2023 Internal Retrospective Unet USA CTs 109 62 47 96.8 95.7 ≤ 2.5 Huhtanen (Model 1) 2022 Internal Retrospective 2D CNN Finland CTs 204 97 107 86.6 93.5 ≤ 3 Huhtanen (Model 2) 2022 Internal Retrospective 2D CNN Finland CTs 204 97 107 83.5 90.7 ≤ 3 Cheikh 2022 External Retrospective DCNN (Aidoc) France CTs 1202 190 1012 92.6 95.8 NR Ajmera 2022 External Retrospective U-Net India CTs 251 55 196 80.0 74 ≤ 2 Buls 2021 External Retrospective DCNN (Aidoc) Belgium CTs 448 83 365 73.0 95 0.5–3.0 Liu 2019 External Retrospective UNet China CTs 62 22 40 100 62.5 NR Liu 2020 External Retrospective CNN, U-Net China CTs 288 186 102 94.6 76.5 0.625–1 Ma 2022 Internal Retrospective 3D CNN, TCN USA CTs 1000 313 687 86.0 85 1 Zhang (ft) 2022 External Retrospective DCNN (ft) China CTs 410 137 273 100 77.29 0.625–1 Zhang (fs) 2022 External Retrospective DCNN (fs) China CTs 410 137 273 100 75.82 0.625–1 Maizlin 2007 External Retrospective CAD Canada CTs 104 15 89 53.3 77.5 1.25 Schmuelling 2021 External Retrospective DL Switzerland CTs 1770 1452 318 79.6 95 1 Weikert 2020 External Retrospective 3D CNN Switzerland CTs 1465 232 1233 92.7 95.5 1 Ebrahimian (optimal quality CT) 2022 External Retrospective 3D CNN USA CTs 197 78 119 96.0 92 1.25 or 1 Ebrahimian (suboptimal quality CT) 2022 External Retrospective 3D CNN USA CTs 133 19 114 100 89 1.25 or 1 Open in a new tab CAD computer aided detection, PENet phenotype encoding network, CNN convolutional neural networks, TCN temporal convolutional network, DL deep learning, mm millimeters, NR not reported Pooled Diagnostic Test Accuracy As there were few data records that did not mention the type of CAD used (whether AI-based or not), three different analyses were performed to determine the pooled accuracy of (1) CAD and AI-based algorithms combined (total), (2) studies explicitly mentioning AI algorithms, and (3) studies explicitly mentioning AI algorithms on external populations (Fig. 2 ). Fig. 2. Open in a new tab Random effects models were used to estimate pooled sensitivity and specificity, with the I 2 index indicating heterogeneity. Forest plot of pooled sensitivity ( a ) and specificity ( b ) of all algorithms or populations included in this study ( I 2 = 87% and 99%, respectively); c , d Forest plot of pooled sensitivity ( c ) and specificity ( d ) of only AI algorithms included in this study ( I 2 = 87% and 95%, respectively); e , f Forest plot of pooled sensitivity ( e ) and specificity ( f ) of only AI algorithms run on an external population ( I . 2 = 87% and 95%, respectively) The total pooled sensitivity and specificity of 26 algorithms/population included in this study were 91.5 (CI 95%; 85.5–95.2) and 84.3 (CI 95% 74.9–0.90.6), respectively. In the univariate approach, the highest sensitivity was for AI algorithms on external populations at 95.4% (CI 95%; 88.8%; 98.2%), and the specificity was similar for AI algorithms on external populations and studies explicitly mentioning AI algorithms at 89.9%. The highest AUC in sROC was for AI algorithms on external populations, with an area under the curve of 0.961; this value was 0.923 for total and 0.951 for studies explicitly mentioning AI algorithms (Table 2 ; Fig. 3 ). Table 2. Meta-analysis based of different group on studies included Group Number of algorithms/populations included Sensitivity I 2 sensitivity Specificity I 2 specificity sROC AUC* Total 26 0.915 (0.855; 0.952) 86.9% 0.843 (0.749; 0.906) 98.7% 0.923 AI algorithms mentioned article 20 0.927 (0.863; 0.962) 87.3% 0.899 (0.850; 0.932) 94.8% 0.951 AI algorithms on external population 14 0.954 (0.888; 0.982) 85.1% 0.899 (0.830; 0.942) 96.3% 0.961 Open in a new tab sROC summary receiver operating characteristic, AUC area under the curve Fig. 3. Open in a new tab Summary receiver operating characteristic (sROC) curves: a for all algorithms or populations included in this study (AUC 0.923, I 2 = 31.2%), b for AI algorithms only (AUC 0.961, I 2 = 47.4%), and c for AI algorithms applied to external population Subgroup Analysis To find the source of heterogeneity, subgroup analysis was performed based on geographical area of the article, year of publication, and slice thickness of CT scan (Table 3 ). Table 3. Subgroup meta-analysis on studies Subgroup analysis Number of algorithms/populations Pooled sensitivity (95% CI) Sensitivity I 2 Pooled specificity (95% CI) Specificity I 2 Geographical area Europe 8 0.881 (0.791; 0.935) 91.0% 0.909 (0.742; 0.972] 99.5% North America 13 0.878 (0.779; 0.936) 86.8% 0.823 (0.676; 0.912) 94.8% Asia 5 0.995 (0.806; 1.000) 60% 0.753 (0.724; 0.781) 6.9% Slice thickness Only less than 1 mm 9 0.955 (0.831; 0.989) 89.4% 0.773 (0.549; 0.905) 99.4% less than and more than 1 mm 12 0.842 (0.746; 0.906) 86.3% 0.885 (0.841; 0.917) 84.9% Year of publication Before 2021 11 0.874 (0.761; 0.938) 89.1% 0.722 (0.528; 0.858) 96.8% After 2021 15 0.936 (0.861; 0.972) 85.5 0.899 (0.816; 0.947) 99.1% Open in a new tab mm millimeters The geographical area was divided into three regions: Europe, North America (USA and Canada), and Asia. In this analysis, the heterogeneity decreased in studies conducted in Asia, with I 2 values for sensitivity and specificity at 60% and 6.9%, respectively. In the slice thickness subgroup analysis, it was divided into two groups: 1—only less than 1 mm, 2—less and more than 1 mm (articles used higher than 1 mm for all or some of their sample size). This subgroup analysis did not show any reduction of heterogeneity. Finally, the study based on the year of publication was categorized into two groups: before and after 2021, but this subgroup analysis also did not show any reduction of heterogeneity. In the next step, the joint model was used to evaluate differences in heterogeneity and diagnostic performance across groups based on the AI algorithm architecture. This analysis showed that heterogeneity ( I 2 ) was highest for the CAD-based and pure DL groups ( I 2 = 99), while the specialized CNN group had lower heterogeneity ( I 2 = 88). The highest AUC values (0.94) were observed in the pure DL and specialized CNN groups, whereas the CAD-based group had the lowest AUC (0.91) (Table 4 ). Table 4. Subgroup meta-analysis based on AI algorithm architecture Subgroup analysis Number of studies Pooled sensitivity (95% CI) Pooled specificity (95% CI) AUC Joint I 2 Classic and advanced CAD 9 0.89 (0.75, 0.95) 0.78 (0.43, 0.94) 0.91 (0.89, 0.94) 99% Pure DL 9 0.95 (0.78, 0.99) 0.88 (0.82, 0.92) 0.94 (0.91, 0.96) 99% Specialized CNN 8 0.90 (0.85, 0.94) 0.86 (0.78, 0.91) 0.94 (0.92, 0.96) 88% Open in a new tab AI artificial intelligence, DL deep learning, CNN convolutional neural network Methodological Quality and Publication Bias In general, studies showed a low-to-moderate risk of bias. The overall risk of bias regarding patient selection was low in flow and timing, reference standard, and index test; however, regarding patients’ selection, this risk was moderate. The applicability of the included study was low in all domains (Fig. 4 ). Also, a funnel plot was drawn, Deeks’ plot and Egger’s and Begg’s tests were performed (supplementary file). Fig. 4. Open in a new tab Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS) for included study for evaluation of risk of bias and applicability Discussion In the present study, we performed a systematic review and meta-analysis to evaluate the performance of different AI algorithms in detecting PE on CTPA. We found that the AI algorithms applied to an external population, the closest environment to the clinical setting, achieved an average sensitivity of 95.4% and a specificity of 89.9% for the detection of PE. Also, when the results were grouped based on the geographical area, the heterogeneity decreased significantly in studies originating from Asia. Based on the previous studies, there was a varying range of sensitivity (53 to 100%) and specificity (14 to 95%) for detection of PE utilizing AI on CTPA. The results of this study demonstrate that using AI algorithms can be promising in the detection of PE on CTPA in a clinical setting. The high sensitivity and specificity of AI for PE detection as demonstrated by the present study require significant attention. AI systems with strong sensitivity and specificity can assist clinicians by enhancing diagnostic precision and reliability, especially in critical situations such as PE [ 21 ]. High sensitivity ensures that most patients are accurately diagnosed, minimizing false negatives and decreasing the risk of missed diagnoses that could postpone crucial treatments—including anticoagulation for PE—and possibly lead to significant injury or death [ 22 , 23 ]. In addition, high specificity effectively excludes individuals without the disease, minimizing false positives and avoiding unnecessary procedures, including unwarranted anticoagulation, which has associated risks of bleeding and other complications [ 23 , 24 ]. A 5% gain in specificity for AI-assisted PE detection in busy emergency departments serving 10,000 patients annually could reduce false positives by 500 cases, reducing unnecessary diagnostic imaging and interventions. This enhances workflow efficiency, reduces patient risk from excessive testing, and lowers clinician alarm fatigue, thereby improving overall care quality in urgent situations [ 25 – 27 ]. AI algorithms have revolutionized the interpretation of medical images, including the detection of PE as well as reducing the time for the diagnosis of PE (Fig. 5 ) [ 28 – 30 ]. However, there are limitations in applying AI algorithms for PE detection on CTPA. A major concern is the requirement for extensive annotated datasets to train the models. Annotated datasets are diverse in size and type, making the process challenging. These limitations lower the applicability of the AI algorithms and, in most cases, necessitate immediate communication with the physician. Furthermore, FP findings may be time-consuming for the radiologists and postpone the workup of patients [ 31 – 35 ]. Fig. 5. Open in a new tab Workflow illustrating artificial intelligence-assisted imaging interpretation for pulmonary embolism on CT scans In a previous systematic review and meta-analysis conducted in 2021 by Soffer et al., five studies were included, and the pooled sensitivity and specificity were 0.88 (95% CI 0.803–0.927) and 0.86 (95% CI 0.756–0.924), respectively. Our study included a large sample size and analyzed 21 articles encompassing more than 26 models/populations. We demonstrated an increased pooled sensitivity and specificity resulted from recent studies included in our analysis [ 36 ]. In addition to demonstrating the high accuracy of AI algorithms for detecting PE, this study revealed that geographical area contributes to heterogeneity. Although many included studies did not comprehensively report the racial composition of their study populations, this analysis suggests that differences in geographical regions may contribute significantly to heterogeneity in imaging data and AI performance. When algorithms are developed using racially or geographically homogeneous datasets, they may underperform in underrepresented populations, raising concerns about diagnostic inequity and algorithmic bias. However, testing AI algorithm performance using external validation datasets from diverse populations and geographic regions, as demonstrated in several original research articles, can support this theory and emphasize the importance of generalizability in clinical AI applications. Finally, besides traditional univariate analysis, we employed bivariate analysis to enhance accuracy by considering sensitivity and specificity together. Various AI algorithms were used for detection of PE on CPTA. CNNs are fundamental AI tools designed to identify patterns through layers of pooling and convolutions while DCNNs are an advanced form of CNNs, consisting of more layers and more complex architecture. DCNNs can extract higher levels of features from imaging and improve the accuracy and precision of detecting emboli within the pulmonary arteries. 3D-DCNN, unlike 2D versions, can process volumetric data and prepare them for analysis. Additionally, 3D-DCNN can detect PE more accurately and reduce false positives. CNN-PENet is another AI algorithm designed for PE detection and advanced feature extraction and conducts multi-scale analysis, enhancing the ability to detect PE at different locations throughout the pulmonary vasculature [ 6 , 37 , 38 ]. There are various challenges in implementing AI in clinical setting. One major concern is ensuring consistent performance of AI across various PE types and clinical situations. Although deep learning systems show high sensitivity and specificity for central and lobar PEs, their effectiveness frequently diminishes for peripheral, segmental, subsegmental, or chronic emboli, leading to inconsistencies in diagnostic reliability among different lesion types [ 39 ]. In addition, algorithm outputs should be customized to clinical thresholds that optimize sensitivity while minimizing false positives to prevent alert fatigue—excessive false alarms can interfere with radiologists and diminish trust. Ultimately, strong, forward-looking validation in diverse clinical settings, along with feedback mechanisms to assess scan coverage and performance after implementation, is crucial for guaranteeing safety and effectiveness [ 39 , 40 ]. Moreover, variability in scanners, populations, and annotations leads to dataset shift in multi-center imaging, which greatly hinders the generalization of AI for detecting PE. In the absence of extensive, varied datasets and strong external validation, models face the danger of inadequate real-world outcomes, highlighting the essential requirement for standardized protocols and improved shift-mitigation techniques in the implementation of clinical AI [ 25 , 41 , 42 ]. Recently, some of these algorithms are being used commercially by institutions to help radiologists in the detection of PE. For instance, Aidoc is an example of an AI-powered solution developed to assist in the detection of PE on CTPA. It uses advanced deep learning algorithms to ensure real-time analysis of imaging data, notifying radiologists to prioritize patients with a high probability of PE, thereby decreasing diagnosis time, thus simplifying workflow and lowering diagnostic delays [ 6 , 43 ]. In addition, a missed diagnosis of PE by AI may place healthcare providers at risk of medicolegal liability, especially if a deviation from standard care or insufficient supervision happens, highlighting the need for well-defined legal structures and collective accountability among clinicians, institutions, and AI developers to reduce risks and protect patient safety. Finally, the integration of AI for PE detection raises important ethical and clinical considerations, including the need for thorough ethical review of training data, obtaining informed patient consent for data use, and ensuring robust data privacy protection. Safeguarding patient confidentiality through strict adherence to privacy regulations and transparent data handling practices is essential to maintaining trust and upholding ethical standards in the development and deployment of AI algorithms in clinical settings. This can be particularly challenging, as many applications are not open-source and always carry a risk of data breaches. Limitations Despite our invaluable findings, this investigation suffers from some limitations. The patients included in this study had varying locations of PE, underwent different types of CT scans device and protocols, and had different sizes of PE, which may lead to inconsistencies. Although the random-effect model was utilized for sensitivity and specificity meta-analysis to mitigate the potential effect of high heterogeneity on the results, it could not completely eliminate this issue. Furthermore, the absence of unpublished studies or abstracts/articles that could not be included in this study due to insufficient data might introduce publication bias like other meta-analysis. In conclusion, AI algorithms in the detection of PE on CTPA is promising in assisting radiologists with acceptable accuracy and high sensitivity and specificity. However, more studies are required to validate AI algorithms as an asset in combating PE, leading to enhancing patient care, and reducing morbidity and mortality with PE. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 517 KB) (517.4KB, docx) Author Contribution Seyed Ali Nabipoorashrafi (SAN): contributed to conceptualization, data extraction, statistical analysis, and drafting the manuscript. Arsalan Seyedi (AS): participated in study design, methodology development, and data extraction. Razman Arabzadeh Bahri (RAB): assisted with literature screening, data validation, and interpretation of findings. Amirhossein Yadegar (AY): involved in data extraction, quality assessment, and visualization of results. Mostafa Shomal-Zadeh (MSZ): conducted data extraction and contributed to figure preparation and manuscript editing. Fatemeh Mohammadi (FM): contributed to literature screening, data entry, and manuscript review. Samira Amin Afshari (SAA): participated in methodological quality assessment and data cross-checking. Negar Firoozeh (NF): reviewed the manuscript critically for important intellectual content and assisted with revisions. Navida Noroozzadeh (NN): performed statistical analysis and prepared visual figures and plots. Farbod Khosravi (FK): supervised the systematic review process, contributed to study design and drafting the manuscript, and edited the final manuscript. Sanaz Asadian (SA): provided critical review, clinical insight on radiological aspects, and oversight of the reporting quality. Hamid Chalian (HC): principal investigator; led project conceptualization, supervised all stages of the study, and critically revised the final manuscript. Funding Hamid Chalian and Seyed Ali Nabipoorashrafi are recipients of GE Healthcare Research Grant. Declarations Ethics Approval This study is a systematic review and meta-analysis of published data and did not involve human participants or animal subjects. Therefore, ethics approval was not required. Consent to Participate Not applicable. Consent for Publication Not applicable. Conflict of interest The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Konstantinides SV, Meyer G, Becattini C, et al. 2019 ESC Guidelines for the diagnosis and management of acute pulmonary embolism developed in collaboration with the European Respiratory Society (ERS): The Task Force for the diagnosis and management of acute pulmonary embolism of the European Society of Cardiology (ESC). European Heart Journal . 2019;41(4):543-603. 10.1093/eurheartj/ehz405 [ DOI ] [ PubMed ] [ Google Scholar ] 2. Baumgartner C, Tritschler T. Clinical significance of subsegmental pulmonary embolism: An ongoing controversy. Research and Practice in Thrombosis and Haemostasis . 2021;5(1):14-16. 10.1002/rth2.12464 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Raja AS, Greenberg JO, Qaseem A, Denberg TD, Fitterman N, Schuur JD. Evaluation of Patients With Suspected Acute Pulmonary Embolism: Best Practice Advice From the Clinical Guidelines Committee of the American College of Physicians. Ann Intern Med . Nov 3 2015;163(9):701-11. 10.7326/m14-1772 [ DOI ] [ PubMed ] [ Google Scholar ] 4. Paul JD, Cifu AS. Management of Acute Pulmonary Embolism. Jama . Aug 11 2020;324(6):597-598. 10.1001/jama.2020.3905 [ DOI ] [ PubMed ] [ Google Scholar ] 5. Li L, Peng M, Zou Y, Li Y, Qiao P. The promise and limitations of artificial intelligence in CTPA-based pulmonary embolism detection. Front Med (Lausanne) . 2025;12:1514931. 10.3389/fmed.2025.1514931 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Cheikh AB, Gorincour G, Nivet H, et al. How artificial intelligence improves radiological interpretation in suspected pulmonary embolism. Eur Radiol . Sep 2022;32(9):5831-5842. 10.1007/s00330-022-08645-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Tajbakhsh N, Shin JY, Gurudu SR, et al. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? IEEE Trans Med Imaging . May 2016;35(5):1299-1312. 10.1109/tmi.2016.2535302 [ DOI ] [ PubMed ] [ Google Scholar ] 8. Ajmera P, Kharat A, Seth J, et al. A deep learning approach for automated diagnosis of pulmonary embolism on computed tomographic pulmonary angiography. BMC Med Imaging . Nov 11 2022;22(1):195. 10.1186/s12880-022-00916-0 [ DOI ] [ PMC free article ] [ PubMed ] 9. Bortsova G, González-Gonzalo C, Wetstein SC, et al. Adversarial attack vulnerability of medical image analysis systems: Unexplored factors. Medical Image Analysis . 2021/10/01/ 2021;73:102141. 10.1016/j.media.2021.102141 [ DOI ] [ PubMed ] 10. Wang W, Wildgruber M, Wang Y. Attacking medical images with minimal noise: exploiting vulnerabilities in medical deep-learning systems. Quantitative Imaging in Medicine and Surgery . 2024;14(12):9374-9384. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Anthimopoulos M, Christodoulidis S, Ebner L, Christe A, Mougiakakou S. Lung Pattern Classification for Interstitial Lung Diseases Using a Deep Convolutional Neural Network. IEEE Trans Med Imaging . May 2016;35(5):1207-1216. 10.1109/tmi.2016.2535865 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Jacob A, Doan B, Granville C. Abstract No. 364 <strong>Man and machine: implications of an artificial intelligence based algorithm in the detection of pulmonary embolism on computed tomography angiography</strong>. Journal of Vascular and Interventional Radiology . 2022;33(6):S166-S167. 10.1016/j.jvir.2022.03.445 13. Baxa J, Ferda J, Bednářová A, et al. Automated detection (CAD) in CT diagnostics of pulmonary embolism: Possible benefit for junior radiologist? journal article. Czech Radiology . 2008;62(3):270-276. [ Google Scholar ] 14. Huhtanen H, Nyman M, Mohsen T, Virkki A, Karlsson A, Hirvonen J. Automated detection of pulmonary embolism from CT-angiograms using deep learning. BMC Med Imaging . Mar 14 2022;22(1):43. 10.1186/s12880-022-00763-z [ DOI ] [ PMC free article ] [ PubMed ] 15. Zhang H, Cheng Y, Chen Z, et al. Clot burden of acute pulmonary thromboembolism: comparison of two deep learning algorithms, Qanadli score, and Mastora score. Quant Imaging Med Surg . 2022;12(1):66-79. 10.21037/qims-21-140 [ DOI ] [ PMC free article ] [ PubMed ] 16. Grenier PA, Ayobi A, Quenet S, et al. Deep Learning-Based Algorithm for Automatic Detection of Pulmonary Embolism in Chest CT Angiograms. Diagnostics . 2023;13(7):1324. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Huang S-C, Pareek A, Zamanian R, Banerjee I, Lungren MP. Multimodal fusion with deep neural networks for leveraging CT imaging and electronic health record: a case-study in pulmonary embolism detection. Scientific Reports . 2020/12/17 2020;10(1):22147. 10.1038/s41598-020-78888-w [ DOI ] [ PMC free article ] [ PubMed ] 18. Ebrahimian S, Digumarthy SR, Homayounieh F, Bizzo BC, Dreyer KJ, Kalra MK. Predictive values of AI-based triage model in suboptimal CT pulmonary angiography. Clin Imaging . Jun 2022;86:25-30. 10.1016/j.clinimag.2022.03.011 [ DOI ] [ PubMed ] [ Google Scholar ] 19. Liu W, Liu M. Artificial Intelligence in Assessment of Acute Pulmonary Embolism on Computed Tomographic Pulmonary Angiography. D28 CLOT ME A RIVER: LATEST DEVELOPMENTS IN DIAGNOSIS AND MANAGEMENT OF PULMONARY EMBOLISMS AND CTEPH . A6051-A6051. 10.1164/ajrccmconference.2019.199.1_meetingabstracts.a6051 20. Schoepf UJ, Schneider AC, Das M, Wood SA, Cheema JI, Costello P. Pulmonary embolism: computer-aided detection at multidetector row spiral computed tomography. J Thorac Imaging . Nov 2007;22(4):319-23. 10.1097/RTI.0b013e31815842a9 [ DOI ] [ PubMed ] [ Google Scholar ] 21. Weikert T, Winkel DJ, Bremerich J, et al. Automated detection of pulmonary embolism in CT pulmonary angiograms using an AI-powered algorithm. European Radiology . 2020/12/01 2020;30(12):6545–6553. 10.1007/s00330-020-06998-0 [ DOI ] [ PubMed ] 22. Vyas V, Sankari A, Goyal A. Acute pulmonary embolism. StatPearls [Internet] . StatPearls Publishing; 2024. [ PubMed ] 23. Hutchinson BD, Navin P, Marom EM, Truong MT, Bruzzi JF. Overdiagnosis of pulmonary embolism by pulmonary CT angiography. American journal of roentgenology . 2015;205(2):271-277. [ DOI ] [ PubMed ] [ Google Scholar ] 24. den Exter PL, van Es J, Erkens PMG, et al. Impact of Delay in Clinical Presentation on the Diagnostic Management and Prognosis of Patients with Suspected Pulmonary Embolism. American Journal of Respiratory and Critical Care Medicine . 2013;187(12):1369-1373. 10.1164/rccm.201212-2219OC [ DOI ] [ PubMed ] [ Google Scholar ] 25. Li L, Peng M, Zou Y, Li Y, Qiao P. The promise and limitations of artificial intelligence in CTPA-based pulmonary embolism detection. Review. Frontiers in Medicine . 2025-March-19 2025;Volume 12 - 202510.3389/fmed.2025.1514931 [ DOI ] [ PMC free article ] [ PubMed ] 26. Soffer S, Klang E, Shimon O, et al. Deep learning for pulmonary embolism detection on computed tomography pulmonary angiogram: a systematic review and meta-analysis. Scientific Reports . 2021/08/04 2021;11(1):15814. 10.1038/s41598-021-95249-3 [ DOI ] [ PMC free article ] [ PubMed ] 27. Langius-Wiffen E, de Jong PA, Hoesein FAM, et al. Retrospective batch analysis to evaluate the diagnostic accuracy of a clinically deployed AI algorithm for the detection of acute pulmonary embolism on CTPA. Insights into Imaging . 2023;14(1):102. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 28. Rothenberg SA, Savage CH, Abou Elkassem A, et al. Prospective Evaluation of AI Triage of Pulmonary Emboli on CT Pulmonary Angiograms. Radiology . Oct 2023;309(1):e230702. 10.1148/radiol.230702 [ DOI ] [ PubMed ] [ Google Scholar ] 29. Pinto-Coelho L. How Artificial Intelligence Is Shaping Medical Imaging Technology: A Survey of Innovations and Applications. Bioengineering (Basel) . Dec 18 2023;10(12)10.3390/bioengineering10121435 [ DOI ] [ PMC free article ] [ PubMed ] 30. Khalifa M, Albadawy M. AI in diagnostic imaging: revolutionising accuracy and efficiency. Computer Methods and programs in biomedicine update . 2024;5:100146. [ Google Scholar ] 31. Zhou C, Chan HP, Sahiner B, et al. Computer-aided detection of pulmonary embolism in computed tomographic pulmonary angiography (CTPA): performance evaluation with independent data sets. Med Phys . Aug 2009;36(8):3385-96. 10.1118/1.3157102 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Shin HC, Roth HR, Gao M, et al. Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Trans Med Imaging . May 2016;35(5):1285-98. 10.1109/tmi.2016.2528162 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Sharma S, Lucas CD. Increasing use of CTPA for the investigation of suspected pulmonary embolism. Postgrad Med . Mar 2017;129(2):193-197. 10.1080/00325481.2017.1281084 [ DOI ] [ PubMed ] [ Google Scholar ] 34. Buhmann S, Herzog P, Liang J, et al. Clinical evaluation of a computer-aided diagnosis (CAD) prototype for the detection of pulmonary embolism. Acad Radiol . Jun 2007;14(6):651-8. 10.1016/j.acra.2007.02.007 [ DOI ] [ PubMed ] [ Google Scholar ] 35. Masoudi M, Pourreza HR, Saadatmand-Tarzjan M, Eftekhari N, Zargar FS, Rad MP. A new dataset of computed-tomography angiography images for computer-aided detection of pulmonary embolism. Sci Data . Sep 4 2018;5:180180. 10.1038/sdata.2018.180 [ DOI ] [ PMC free article ] [ PubMed ] 36. Soffer S, Klang E, Shimon O, et al. Deep learning for pulmonary embolism detection on computed tomography pulmonary angiogram: a systematic review and meta-analysis. Sci Rep . Aug 4 2021;11(1):15814. 10.1038/s41598-021-95249-3 [ DOI ] [ PMC free article ] [ PubMed ] 37. Huang SC, Kothari T, Banerjee I, et al. PENet-a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging. NPJ Digit Med . 2020;3:61. 10.1038/s41746-020-0266-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 38. Vainio T, Mäkelä T, Savolainen S, Kangasniemi M. Performance of a 3D convolutional neural network in the detection of hypoperfusion at CT pulmonary angiography in patients with chronic pulmonary embolism: a feasibility study. Eur Radiol Exp . Sep 24 2021;5(1):45. 10.1186/s41747-021-00235-z [ DOI ] [ PMC free article ] [ PubMed ] 39. Zsarnoczay E, Rapaka S, Schoepf UJ, et al. Accuracy of a deep neural network for automated pulmonary embolism detection on dedicated CT pulmonary angiograms. European Journal of Radiology . 2025;18710.1016/j.ejrad.2025.112077 [ DOI ] [ PubMed ] 40. Blezek DJ, Olson-Williams L, Missert A, Korfiatis P. AI Integration in the Clinical Workflow. Journal of Digital Imaging . 2021/12/01 2021;34(6):1435–1446. 10.1007/s10278-021-00525-3 [ DOI ] [ PMC free article ] [ PubMed ] 41. Kushol R, Wilman AH, Kalra S, Yang Y-H. DSMRI: Domain Shift Analyzer for Multi-Center MRI Datasets. Diagnostics . 2023;13(18):2947. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Guo X, Gichoya JW, Trivedi H, Purkayastha S, Banerjee I. Medshift: automated identification of shift data for medical image dataset curation. IEEE journal of biomedical and health informatics . 2023;27(8):3936-3947. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Langius-Wiffen E, de Jong PA, Hoesein FAM, et al. Retrospective batch analysis to evaluate the diagnostic accuracy of a clinically deployed AI algorithm for the detection of acute pulmonary embolism on CTPA. Insights Imaging . Jun 6 2023;14(1):102. 10.1186/s13244-023-01454-1 [ DOI ] [ PMC free article ] [ PubMed ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary file1 (DOCX 517 KB) (517.4KB, docx) Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer ACTIONS View on publisher site PDF (2.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Artificial intelligence in computed tomography imaging for pulmonary embolism: a narrative review from computed tomography pulmonary angiography to non-contrast computed tomography.#662689
• ctDNA Mutations and Methylation Status for Early Detection of Lung Cancer in Patients With Suspicious Lung Nodules#832488
• Exploring Novel Biomarkers for Emphysema Detection#1115185
• Evaluating the Real World Performance of an AI Based Lung Nodule Detection Tool#895003
• First steps in clinical implementation of a lung nodule classification method by artificial intelligence.#430752