Pulmonary artery sarcoma (PAS) is a rare and highly lethal malignancy that arises directly from the wall of the pulmonary artery. Its most clinically dangerous property is that it produces filling defects inside the pulmonary arteries on CT pulmonary angiography (CTPA) scans, the same imaging appearance as pulmonary thromboembolism (PTE). Because PTE is far more common, PAS patients are routinely misdiagnosed as having blood clots and placed on anticoagulation therapy. This delay to correct diagnosis is catastrophic: PAS is best treated with surgical resection of the pulmonary artery, and every week spent on inappropriate anticoagulation is a week the tumor is progressing.
Why differentiation is so hard: Both PTE and PAS fill the lumen of the pulmonary arteries with material that blocks contrast on CTPA. Advanced imaging modalities, specifically 18F-FDG PET/CT and MRI, can reliably distinguish the two conditions based on metabolic activity and soft-tissue signal characteristics, but these are rarely used as first-line investigations in patients presenting with suspected pulmonary vascular disease. CTPA remains the standard initial workup, and radiologists must rely on subtle morphological clues, such as vessel expansion, extraluminal extension, and the "wall eclipsing sign," that are present in some but not all PAS cases.
The literature gap: The authors conducted a systematic PubMed and Google Scholar search through November 2024 using terms including "pulmonary artery sarcoma," "pulmonary embolism," "computed tomography pulmonary angiography," and "deep learning." Their search confirmed that prior studies addressing PAS versus PTE differentiation were almost entirely case reports, small observational series, or systematic reviews of imaging features. No deep learning model for this specific differential diagnosis existed. This study from the CHOICE (CHinese pulmOnary embolism multimodality Imaging artifiCial intelligencE) consortium aimed to fill that gap by building and validating the first such model, named PVDNet.
A further complication is that PTE itself encompasses two clinically distinct entities with different treatment pathways: acute pulmonary thromboembolism (APE), which is managed with immediate anticoagulation or thrombolysis, and chronic pulmonary thromboembolism (CPE), which may require balloon pulmonary angioplasty (BPA) or pulmonary endarterectomy (PEA). The three-way classification task of distinguishing APE, CPE, and PAS simultaneously is therefore clinically essential and considerably harder than a simple binary decision.
This study drew its imaging data from the prospective CHOICE study (NCT06526468), a registered multicenter observational study. The retrospective analysis of CTPA datasets was structured around a deliberate non-random splitting strategy designed to simulate real-world model deployment rather than artificially inflated cross-validation performance. The training set comprised scans collected at three centers between January 2015 and December 2019, while the internal test set used scans from those same three centers but from a later period, January 2020 to December 2023. This temporal separation means the model was trained on historically older data and tested on more recent data, mimicking the situation where a deployed model encounters patients scanned after its training window closed.
Patient composition: A total of 1,587 consecutive CTPA studies were initially identified from 15 participating centers. After applying exclusion criteria (poor image quality, absence of DICOM format, lost to follow-up, or diagnoses of other rare pulmonary arterial conditions such as fibrinous mediastinitis, Takayasu arteritis, Behcet's disease, hydatid pulmonary embolism, and pulmonary artery lipoma), 952 eligible studies were retained. The final dataset included 470 APE patients (mean age 65.5 +/- 15.2 years), 363 CPE patients (mean age 58.4 +/- 14.0 years), and 119 PAS patients (mean age 51.5 +/- 14.6 years). Notably, PAS patients were significantly younger than the PTE groups (F = 52.109, p < 0.001), which has clinical relevance as younger patients with unexplained pulmonary artery filling defects warrant higher diagnostic suspicion for sarcoma.
Diagnosis confirmation: PAS cases required histopathological confirmation via either endovascular biopsy (56 cases) or surgical pathology (63 cases, representing 52.9% of PAS patients). APE required baseline CTPA-confirmed filling defects plus clinical and laboratory corroboration plus either improvement after at least 3 months of anticoagulation or follow-up imaging showing defect reduction. CPE required persistent exertional symptoms despite at least 3 months of effective anticoagulation, V/Q scintigraphy showing at least 2 mismatched segmental or lobar perfusion defects, and CTPA or pulmonary angiography confirming chronic thromboembolic signs. Ambiguous cases were adjudicated by a multidisciplinary panel. The study adhered to the TRIPOD-AI reporting checklist.
Dataset partitioning: The 590-patient training set generated 390,816 image slices from three centers. The 186-patient internal test set contributed 65,198 slices from the same three centers across the later time window. The external validation set comprised 176 patients from 12 entirely independent centers, generating 60,720 slices. The PAS-to-PTE ratio was approximately 1:7, creating a significant class imbalance that required explicit mitigation during model training.
The PVDNet framework is built on COVNet, a supervised convolutional neural network originally developed for COVID-19 lung classification, repurposed and substantially modified for pulmonary vascular disease differentiation. Its backbone is ResNet-50, a 50-layer residual deep learning model with 3.8 billion trainable parameters. ResNet-50 is particularly suited to medical imaging tasks because its residual connections address the vanishing gradient problem that afflicts very deep networks, enabling effective feature learning from complex volumetric CT data. The ResNet-50 weights were initialized from ImageNet pretraining, with the final classification layer reinitialized using the Kaiming method to adapt the network to this three-class medical imaging problem.
Domain-specific modifications: Three targeted architectural changes were made to adapt the ResNet-50 backbone for pulmonary vascular disease. First, max-pooling-based feature aggregation was introduced to prioritize discriminative spatial patterns and suppress noise, which is particularly important for detecting rare PAS features amid the vastly more common PTE appearance. Second, a fine-grained classification head consisting of a fully connected layer followed by a softmax activation was added to produce probability scores for each of the three classes (APE, CPE, PAS). Third, integrated Grad-CAM (Gradient-weighted Class Activation Mapping) modules were embedded to generate interpretable saliency maps that highlight which image regions drove each classification decision, enabling radiologists to understand and verify the model's reasoning.
Training configuration: The model was trained for 50 epochs using the Adam optimizer with an initial learning rate of 1e-5 and mini-batch size of 1. Stochastic gradient descent with momentum of 0.9 was used to accelerate convergence while avoiding local minima. The shallow ResNet-50 layers (Conv1 through Conv2_x) were frozen to preserve general visual features learned during ImageNet pretraining, while deeper layers (Conv3_x onward) were fine-tuned on CTPA data. Data augmentation was performed via the batchgenerators toolkit. The ReduceLROnPlateau scheduler adjusted learning rate based on tuning set loss. A five-fold cross-validation procedure was used for model development, with the mean F1 score across folds reported as 0.838 +/- 0.051 on the tuning set.
Image preprocessing pipeline: Each CTPA study in DICOM format underwent four preprocessing steps before entering the model. All images were standardized to a window center of 100 HU and window width of 700 HU to equalize contrast across different scanners and centers. A validated DL-based lung segmentation tool (InferRead CT Lung v3.12.3, Infervision Medical Technology) automatically extracted lung masks. The lung region was extracted via a 3D bounding box preserving the pulmonary arteries and heart. Each volume was resized to 192 x 192 x 192 voxels to meet ResNet-50 input requirements, then subjected to min-max normalization to reduce gray-level variation across different scanning parameters.
Class imbalance handling: Given the 7:1 PTE-to-PAS ratio, two strategies were applied. Class-balanced sampling oversampled PAS cases in the training batches so the model encountered PAS examples at a frequency proportionate to its diagnostic importance rather than its statistical rarity. Class-balanced loss replaced the standard cross-entropy loss function, weighting misclassification of rare classes (PAS) more heavily during training. Together these approaches directed model learning toward accurate PAS detection despite its underrepresentation.
In the internal test set of 186 patients from the same three training centers but a later time period, PVDNet achieved strong performance across all three classification categories. For PAS detection, the model achieved an AUC of 0.972 (95% CI: 0.945, 0.994). For APE classification, AUC was 0.902 (95% CI: 0.855, 0.944). For CPE classification, AUC was 0.900 (95% CI: 0.852, 0.946). The overall agreement between PVDNet predictions and ground-truth labels was moderate (Cohen's kappa = 0.687, p < 0.001), with kappa values from 0.4 to 0.75 classified as moderate consistency under the study's pre-specified scale. In binary classification (PAS versus combined PTE), the model achieved an AUC of 0.97, demonstrating that the three-class formulation did not meaningfully compromise performance on the primary clinical question of sarcoma detection.
Interpretability via Grad-CAM: Gradient-weighted Class Activation Maps were generated to visualize model attention for representative APE, CPE, and PAS cases. The heatmaps confirmed that the model focused on the anatomically relevant pulmonary artery filling defects when making its decisions, rather than spurious background features. This alignment between model attention and known radiological diagnostic criteria increases clinician confidence in model outputs and provides a pathway for identifying cases where the model may be attending to the wrong regions.
Per-fold consistency: Five-fold cross-validation during training yielded consistent performance across folds (detailed in Appendix Table S3), confirming that the reported internal test set results were not the product of a particularly favorable data split. The five trained models were evaluated independently on the internal test set rather than ensembled, with their performance metrics averaged to provide a comprehensive assessment. This approach preserves interpretability while providing a robust estimate of expected real-world performance.
The external validation set of 176 patients from 12 institutions not involved in model development is the most clinically meaningful test of PVDNet's generalizability. Four radiologists with different experience levels independently evaluated all external validation cases under strict blinding conditions: they received anonymized full CTPA images without access to clinical history, laboratory findings, prior imaging reports, or final diagnoses. Image datasets were randomized before distribution to eliminate sequence bias. Radiologists were permitted window width and level adjustments consistent with routine clinical workflow but were otherwise constrained to structured reporting templates.
PAS classification performance: PVDNet achieved an AUC of 0.973 (95% CI: 0.951, 0.992) for PAS detection in the external validation set. The senior radiologist specializing in pulmonary vascular diseases with 20 years of experience (SRPV) achieved an AUC of 0.943 (95% CI: 0.873, 0.997). The difference between PVDNet and SRPV was not statistically significant (p = 0.308, DeLong's test), indicating that the model performs at the level of a subspecialty expert. By contrast, PVDNet significantly outperformed all three less experienced radiologists: the mid-career radiologist with 13 years of experience (MRPV) achieved AUC 0.872 (p < 0.050 vs. PVDNet), the junior radiologist with 6 years of experience (JRPV) achieved AUC 0.793 (p < 0.001), and the junior resident with 3 years of experience (JRDS) achieved AUC 0.741 (p < 0.001).
Agreement analysis: Cohen's kappa between PVDNet predictions and ground-truth labels was 0.651 (p < 0.001), representing moderate consistency. This was the highest agreement coefficient among all four readers: SRPV achieved kappa = 0.554, wait, specifically: JRDS kappa = 0.282, JRPV kappa = 0.475, MRPV kappa = 0.554, SRPV kappa = 0.651 was matched by the model at 0.651. The model's agreement level therefore matched the most experienced human expert.
APE and CPE performance: For APE classification in the external validation set, PVDNet (AUC 0.835, 95% CI: 0.775, 0.893) significantly outperformed the junior resident JRDS (AUC 0.686, p < 0.010). For CPE classification, PVDNet achieved AUC 0.809 (95% CI: 0.731, 0.884), which was comparable to JRDS (AUC 0.735, p = 0.230) but did not outperform more experienced readers. The binary PAS-versus-PTE classification AUC on external validation was 0.97, consistent with internal test set findings.
PVDNet's performance profile is not uniform across the three diagnostic categories. The model demonstrates exceptional accuracy for PAS detection in both internal and external sets, consistent performance for APE classification, and noticeably weaker performance for CPE classification, particularly in the external validation set where APE and CPE performance dropped relative to internal benchmarks. This pattern reflects fundamental differences in the complexity of the imaging features that distinguish each condition.
Why PAS is the model's strength: PAS produces several relatively distinctive morphological features on CTPA, including expansion of the involved pulmonary artery beyond what a blood clot alone would cause, the wall eclipsing sign (where tumor attenuates the vessel wall), and extraluminal extension of soft tissue beyond the arterial boundary. While these features are not universally present, they provide enough consistent signal for a deep learning model to learn discriminative representations, especially when trained on what the authors describe as the largest PAS cohort assembled to date (119 pathologically confirmed cases).
Why APE-CPE discrimination is harder: Distinguishing acute from chronic thromboembolism requires integrating information that goes beyond simple morphological appearance. Experienced radiologists factor in secondary changes in the vascular wall structure, bronchovascular relationships, evidence of pulmonary hypertension sequelae, parenchymal changes from chronic hypoperfusion, and temporal context from prior imaging. A single-timepoint CTPA analysis, which is what PVDNet currently uses, cannot capture the dynamic trajectory from acute to chronic phases. The authors explicitly note that future work should develop multitemporal trajectory models using longitudinal CHOICE cohort data to address this specific limitation.
External versus internal performance gap: The drop in APE and CPE performance between the internal test set (AUC approximately 0.90 for both) and the external validation set (AUC approximately 0.83 for APE and 0.81 for CPE) was attributed to two compounding factors. First, the external validation set had a higher proportion of APE patients and a lower proportion of CPE patients relative to the training distribution, creating a mismatch that disadvantages CPE classification. Second, the 12-center external set exhibited greater scanner and protocol heterogeneity than the 3-center internal set. PAS classification was relatively unaffected by these factors, likely because its distinctive morphological features are more salient and less sensitive to subtle contrast or resolution variations than the more nuanced APE-versus-CPE distinctions.
Retrospective design and selection bias: Although the underlying CHOICE study is prospective, PVDNet was developed through retrospective analysis of a subset of the CHOICE data. Retrospective analyses are vulnerable to case selection biases, including the enrichment of PAS cases at referral centers (the study's PAS proportion of 12.5% far exceeds population incidence of this rare tumor). While this enrichment was intentional and necessary to build a robust model for a rare diagnosis, it means that deploying PVDNet in low-prevalence community settings would encounter a very different pretest probability distribution. The authors argue that PVDNet's high specificity in external validation would prevent a clinically significant increase in false positives even at low prevalence, but prospective real-world validation is needed to confirm this.
CTPA-only architecture: PVDNet currently processes only CTPA imaging data. It does not incorporate any clinical metadata, laboratory biomarkers (such as D-dimer, brain natriuretic peptide, or LDH), symptom duration, or prior imaging history. Integration of these multimodal inputs could substantially improve diagnostic capabilities, particularly for the more nuanced APE-CPE differentiation task where clinical context is highly informative. This is a known limitation of purely image-based deep learning models in clinical medicine.
Tumor thrombus not evaluated: Endovascular pulmonary artery involvement by metastatic tumor thrombi (occurring in patients with systemic cancers such as renal cell carcinoma or hepatocellular carcinoma) was not included as a model category due to insufficient pathologically confirmed cases in the dataset. Tumor thrombi can produce CTPA appearances overlapping with both PTE and PAS. As more confirmed tumor thrombus cases accumulate in the CHOICE cohort, extending PVDNet to distinguish this additional entity will be an important research direction.
No human-AI collaboration study: The study compared AI-only versus human-only diagnostic performance but did not study the hybrid paradigm where radiologists interpret scans with AI assistance. This is a clinically important gap, as AI tools in radiology are typically deployed as decision support rather than autonomous diagnostic systems. Whether PVDNet can measurably close the diagnostic gap between junior and senior radiologists when used as a second reader remains an open and practically significant question.
Sex-stratified analysis absent: Given known sex differences in pulmonary vascular physiology and disease presentation, the authors acknowledge that sex-specific variations in imaging features could influence model performance. Subgroup analyses by sex were not feasible in this study due to sample size limitations in certain subgroups, particularly for PAS. Future studies with larger cohorts should systematically assess whether sex-stratified models or sex as an additional input feature improves diagnostic accuracy across populations.
Prospective real-world validation: The authors identify prospective validation as the highest priority next step for PVDNet. A prospective study embedding PVDNet into clinical CTPA reporting workflows at diverse center types, including community hospitals where subspecialty expertise in pulmonary vascular disease is scarce, would generate the evidence needed to establish its clinical utility and define its appropriate role in the diagnostic pathway. Such a study should measure not just AUC metrics but clinically meaningful endpoints: time to correct diagnosis, rate of delayed PAS identification, frequency of inappropriate anticoagulation in PAS patients, and ultimately surgical referral rates and survival outcomes.
Multimodal model development: The current architecture should be extended to incorporate clinical metadata alongside imaging data. Clinically available variables such as D-dimer levels, symptom duration, prior venous thromboembolism history, echocardiographic parameters, and laboratory markers of tumor burden could all theoretically improve discrimination, especially for the APE-CPE boundary case that currently represents PVDNet's weakest performance domain. The CHOICE cohort's multimodality data infrastructure, which includes PET/CT and MRI in a subset of patients, provides an existing foundation for developing these multimodal extensions.
Temporal trajectory modeling: To address the specific challenge of APE-CPE differentiation, the authors propose constructing multitemporal trajectory models from longitudinal CHOICE cohort data. By incorporating serial CTPA examinations and capturing the morphological evolution of filling defects over time, a future model could learn to distinguish acute from chronic thromboembolic patterns with the same dynamic reasoning that experienced radiologists apply in practice. This would require careful temporal data curation and model architectures capable of processing sequences of imaging studies.
Harmonization and scanner generalization: The performance differences observed between the 3-center internal validation and the 12-center external validation highlight the ongoing challenge of cross-scanner generalization in medical imaging AI. Future iterations could apply imaging harmonization tools such as ComBat or deep learning-based domain adaptation methods to reduce inter-scanner feature variability before it enters the classification pipeline. These approaches have shown promise in radiomics and could improve PVDNet's consistency across the diverse scanner ecosystem (Siemens, GE, Philips, Canon, and United Imaging scanners were all represented in the CHOICE cohort).
Clinical impact on underserved settings: PVDNet's most transformative potential lies not in replacing senior subspecialty radiologists who already perform comparably, but in augmenting the capabilities of junior radiologists and general radiologists at centers without pulmonary vascular disease expertise. The study demonstrated that PVDNet significantly outperformed three of four readers including mid-career and junior radiologists for PAS detection. Deployed as a decision-support tool in settings where specialized expertise is unavailable, PVDNet could trigger earlier appropriate referrals to pulmonary vascular disease centers for biopsy and surgical evaluation before disease has progressed beyond resectability.