3D region-growing nnU-Net improves pulmonary embolism detection on CTPA: a dual-cohort validation study

We compared three customized nnU-Net models (A: baseline two-dimensional (2D); B: 2D + region-growing; C: three-dimensional (3D) + region-growing) for automated detection and blood clot volume (BCV) quantification of acute pulmonary embolism (PE) on computed tomography pulmonary angiography (CTPA), and to explore the association between BCV and clinical outcome.

We retrospectively screened 9,715 CTPA examinations (2015‒2024) to develop a dataset of 874 PE-positive and 339 PE-negative cases. A stratified subset (n = 437) with manually refined ground-truth segmentations was used for model training and internal validation. Region-growing in Models B and C included a 5-voxel negative buffer. Internal testing was performed on 776 cases (Humanitas dataset). External testing was performed on the public RSPECT-RSNA dataset. Performance metrics included accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) at zero-clot and for optimized BCV threshold. Correlations between BCV, survival, and major adverse cardiovascular events (MACE) were analyzed.

Model C achieved the highest AUROC on external testing (0.868), outperforming Model A (0.843) and Model B (0.846). On internal testing at ROC-optimized threshold, Model C showed the highest accuracy (85.5%) and AUROC (0.909) compared to Model A (73.4%, 0.784) and Model B (76.0%, 0.816). Model C achieved 83.6% sensitivity and 79.5% accuracy at the zero-clot threshold on external data. BCV was not significantly associated with MACE or survival (p = 0.600).

A locally trained 3D nnU-Net with region-growing demonstrated superior performance and generalizability on external data for automated PE detection on CTPA. However, BCV was not predictive of short-term clinical outcomes.

A locally developed nnU-Net models integrating volumetric 3D segmentation with region-growing offer robust, clinically acceptable performance for the detection of acute pulmonary embolism without the need for ROC-based thresholds.

Our 3D nnU-Net model automates clot detection on CT scans in seconds and shows numerically higher performance than the 2D models. Built on local data, this framework enables institution-specific model training and validation to complement European conformity‒CE-marked tools and assess performance locally. High-sensitivity volumetric quantification reduces missed emboli, paving the way for personalized risk stratification and improved patient outcomes.