Despite the recognized clinical relevance of BCV, its routine use in the standard management of PE remains out of reach because there is no universally available software capable of providing the clot segmentation. The semi-manual segmentation process is time-consuming, operator-dependent, and offers low reproducibility across centers, which limits thrombus volumetry to isolated research protocols.

ClotIA was developed to address this gap, demonstrating that accurate, rapid, and reproducible 3D thrombus quantification is achievable on a large scale. ClotIA achieves a mean DSC of 0.828 after five edits, significantly higher than the baseline (p < 0.001) and more accurate than nnUNet. This precision is well below the typical inter-observer variability of manual segmentation (8–12 mL), making the BCV derived from ClotIA a reproducible, operator-independent biomarker and a reliable data point for future analyses [7].

While ClotIA provides an accessible and objective quantification of the blood clot volume (BCV), it is important to recognize that this biomarker alone does not capture the full complexity of the condition. More specifically, BCV does not account for the anatomical location or spatial distribution of the thrombus, nor does it reflect the patient’s underlying pulmonary vascular reserve.

It is essential to recognize these limitations, as the hemodynamic impact of a given volume remains intrinsically linked to these unmeasured variables and the patient’s specific cardiopulmonary adaptation. In this way, the clinical value of BCV is fully realized when it serves as a complementary tool within a broader multiparametric assessment, ensuring that objective volumetric data are interpreted with consideration of the clinical and hemodynamic context. Previous studies support this integrative approach by showing that combining thrombus volume with cardiac imaging parameters and markers such as troponin I or NT-proBNP significantly improves the prediction of right ventricular dysfunction compared to the use of a single parameter [13, 26, 27], similarly to the integration of these biomarkers into existing scoring frameworks like Qanadli score, allows for better risk stratification [28]. This is precisely where the strength of a foundation model-based approach lies within the inherent limitations of simple volumetry. Unlike traditional AI specialized in specific tasks, FMs benefit from a unified representation space and a deeper “understanding of the image” that allows it to see beyond the 3D extent of the clot, pay “attention” to the surrounding environment, and thus weight its outputs accordingly.

In the future, ClotIA could be integrated into a comprehensive decision-support tool for personalized interventional therapy. Previous studies using standard AI models analyzing one or more imaging parameters are moving in this direction, notably with texture analysis enabling the prediction of short-term prognosis [9], radiomic scores outperforming clinical scores in predicting adverse outcomes [10], and dual-energy CT radiomics enabling independent risk stratification [11, 12].

Several limitations should be considered. ClotIA was evaluated on a cohort of 309 patients using primarily high-quality scans with optimal contrast and minimal artifacts. In real-world emergency situations, suboptimal bolus timing or motion artifacts could affect performance. Furthermore, although we demonstrated a strong correlation between BCV and right ventricular strain, the lack of an external validation cohort and direct correlation with clinical outcomes currently limits its immediate generalizability. Future steps are already underway to address these gaps through external validation studies to confirm generalizability and assess clinical utility in therapeutic decision-making in interventional radiology.