Rethinking Pulmonary Embolism Segmentation: A Study of Current Approaches and Challenges with an Open Weight Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Pulmonary embolism (PE) is a critical condition requiring precise detection.
• Current PE segmentation algorithms face limitations due to small datasets and lack of reproducibility.
• Advancing PE detection relies on robust and comparable algorithm evaluation.

Purpose of the Study:

• To systematically evaluate and compare various PE segmentation algorithms.
• To establish a reproducible baseline for PE segmentation research.
• To identify challenges and guide future research in PE detection.

Main Methods:

• Curated a dataset of 490 CT pulmonary angiography (CTPA) scans with dense annotations.
• Evaluated nine segmentation architectures (CNN and ViT-based) in 2D and 3D.
• Utilized mean Dice similarity coefficient (mDSC) and average symmetric surface distance (ASSD) for evaluation.

Main Results:

• A 3D U-Net with ResNet encoding blocks demonstrated superior PE segmentation performance.
• 3D models consistently outperformed 2D models across all evaluated architectures.
• Model error patterns were consistent across architectures when trained on the same data.
• Distal emboli remain challenging due to task complexity and data scarcity.

Conclusions:

• 3D U-Net architectures are highly effective for pulmonary embolism segmentation.
• Enhanced datasets are crucial for improving distal emboli detection.
• Publicly sharing models and code promotes reproducibility in medical image analysis research.