Meta-analysis of AI-based pulmonary embolism detection: How reliable are deep learning models?

Rationale and objectives

Deep learning (DL)–based methods show promise in detecting pulmonary embolism (PE) on CT pulmonary angiography (CTPA), potentially improving diagnostic accuracy and workflow efficiency. This meta-analysis aimed to (1) determine pooled performance estimates of DL algorithms for PE detection; and (2) compare the diagnostic efficacy of convolutional neural network (CNN)– versus U-Net–based architectures.

Materials and methods

Following PRISMA guidelines, we searched PubMed and EMBASE through April 15, 2025 for English‐language studies (2010–2025) reporting DL models for PE detection with extractable 2 × 2 data or performance metrics. True/false positives and negatives were reconstructed when necessary under an assumed 50 % PE prevalence (with 0.5 continuity correction). We approximated AUROC as the mean of sensitivity and specificity if not directly reported. Sensitivity, specificity, accuracy, PPV and NPV were pooled using a DerSimonian–Laird random-effects model with Freeman-Tukey transformation; AUROC values were combined via a fixed-effect inverse-variance approach. Heterogeneity was assessed by Cochran's Q and I². Subgroup analyses contrasted CNN versus U-Net models.

Results

Twenty-four studies (n = 22,984 patients) met inclusion criteria. Pooled estimates were: AUROC 0.895 (95 % CI: 0.874–0.917), sensitivity 0.894 (0.856–0.923), specificity 0.871 (0.831–0.903), accuracy 0.857 (0.833–0.882), PPV 0.832 (0.794–0.869) and NPV 0.902 (0.874–0.929). Between-study heterogeneity was high (I² ≈ 97 % for sensitivity/specificity). U-Net models exhibited higher sensitivity (0.899 vs 0.893) and CNN models higher specificity (0.926 vs 0.900); subgroup Q‐tests confirmed significant differences for both sensitivity (p = 0.0002) and specificity (p < 0.001).

Conclusions

DL algorithms demonstrate high diagnostic accuracy for PE detection on CTPA, with complementary strengths: U-Net architectures excel in true-positive identification, whereas CNNs yield fewer false positives. However, marked heterogeneity underscores the need for standardized, prospective validation before routine clinical implementation.