Automated Segmentation of Acute Pulmonary Embolism in Cardiac CT Scans Using Deep Learning

Overview

This study evaluates two advanced deep learning architectures, nnU-Net and VT-UNet, for automated segmentation of acute pulmonary embolism (APE) in 3D cardiac CT pulmonary angiography (CTPA) scans. The models were trained and validated on a dataset of 200 volumetric CTPA scans with manual segmentations, demonstrating potential to reduce manual annotation time and improve diagnostic precision.

Background

Acute pulmonary embolism (APE) is a critical cardiovascular condition with mortality rates up to 30%, requiring prompt and accurate diagnosis. Computed tomographic pulmonary angiography (CTPA) is the gold standard for APE diagnosis, where precise segmentation of emboli is essential for risk stratification and treatment decisions. Manual segmentation is labor-intensive and subjective, motivating the development of automated deep learning methods to enhance detection and delineation of emboli in volumetric CTPA images.

Data Highlights

Key Findings

• Manual segmentation of APE in volumetric CTPA is time-consuming, initially requiring about 1 hour per patient, reduced to 15-20 minutes with iterative deep learning-assisted corrections.
• The nnU-Net architecture automatically configures optimized UNet variants for APE segmentation, leveraging dataset-specific heuristics.
• The VT-UNet transformer-based model captures both local and global spatial context via window-based multi-head self-attention, enhancing segmentation accuracy.
• Both models were applied to whole 3D CTPA volumes, addressing limitations of prior 2D or patch-based approaches.
• Automated segmentation provides detailed quantitative information on emboli location, volume, and morphology, critical for patient risk stratification and treatment planning.

Clinical Implications

Automated deep learning segmentation of APE in 3D CTPA scans can significantly reduce radiologist workload and inter-observer variability, enabling faster and more precise diagnosis. Detailed volumetric emboli delineation supports improved risk stratification, guiding appropriate therapeutic decisions such as thrombolytic therapy or anticoagulation management.

Conclusion

The study demonstrates that state-of-the-art deep learning models, including nnU-Net and VT-UNet, effectively segment acute pulmonary embolism in volumetric CTPA images, offering a promising tool to enhance clinical workflows and patient outcomes.

References

1. 1 -- Epidemiology of Acute Pulmonary Embolism
2. 2 -- Mortality Rates in Acute Pulmonary Embolism
3. 3,4 -- CTPA Parameters and Mortality Correlation
4. 5,6 -- Treatment Guidelines Based on Risk Stratification
5. 7 -- Challenges in Manual APE Segmentation
6. 8,9 -- CNN-based APE Classification Studies
7. 10-13 -- Automated APE Detection Approaches
8. 14 -- Liu et al. U-shaped Network for 2D APE Segmentation
9. 15 -- Pu et al. Two-step Pulmonary Artery and APE Segmentation
10. 16 -- Chen et al. Swin Transformer-based APE Segmentation
11. 17 -- nnU-Net Architecture
12. 18 -- VT-UNet Transformer-based Segmentation Model
13. 19 -- Original UNet Architecture