nnU-Net-based deep-learning for pulmonary embolism: detection, clot volume quantification, and severity correlation...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Diagnostic Imaging
• Pulmonary Embolism Detection

Background:

• CT pulmonary angiography is the standard for diagnosing pulmonary embolism (PE).
• Increasing demand necessitates efficient diagnostic tools like deep learning (DL) algorithms.
• The nnU-Net framework simplifies DL model development for medical imaging.

Purpose of the Study:

• To evaluate a locally developed nnU-Net algorithm for PE detection.
• To assess the algorithm's ability to measure blood clot volume (BCV).
• To correlate BCV with right ventricle (RV) overload in PE patients.

Main Methods:

• Trained nnU-Net on the RSPECT dataset (205 PE, 340 negative cases).
• Tested on 6573 exams, analyzing PE characteristics and RV overload.
• Extracted BCV and used ROC curves and logistic regression for validation.

Main Results:

• Median BCV was 1 μL in negatives, 345 μL in PE-positive, and 7,378 μL in central PEs.
• Significant correlation found between BCV, PE presence, central PE, and RV/LV ratio (p < 0.0001).
• AUC for PE detection was 0.865 (83% accuracy), central PE 0.937 (91% accuracy), and RV overload 0.848 (79% accuracy).

Conclusions:

• nnU-Net accurately detects PE, especially central PE.
• BCV is a reliable metric for automated severity stratification and case prioritization.
• nnU-Net provides a user-friendly DL solution for PE detection and severity assessment.