AI-Enhanced Deep Learning Framework for Pulmonary Embolism Detection in CT Angiography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Deep Learning for Medical Diagnosis

Background:

• Pulmonary embolism (PE) detection on CT pulmonary angiography (CTPA) is critical but challenging for small, low-contrast clots.
• Accurate segmentation of distal emboli is essential for effective PE diagnosis and management.

Purpose of the Study:

• To benchmark fully convolutional network (FCN) backbones for PE segmentation.
• To introduce and evaluate Consensus Intersection-Optimized Fusion (CIOF) for improved PE detection accuracy.

Main Methods:

• Ten FCN backbones were benchmarked, and CIOF was developed as a K-of-M pixel-wise mask fusion technique.
• The FUMPE cohort (35 patients) was used for training (18) and testing (17) with patient-based random splits.
• Segmentation performance was evaluated using Intersection over Union (IoU), Dice coefficient, false negative/positive rates (FNR/FPR), and latency.

Main Results:

• CIOF achieved superior overall performance (mean IoU 0.569, Dice 0.691) compared to single backbones.
• CIOF demonstrated significant improvements in segmenting small and subsegmental emboli across different clot burdens.
• The strongest single backbone, Inception-ResNetV2 + SGDM, showed competitive performance (IoU 0.530, Dice 0.648) with lower latency.

Conclusions:

• CIOF offers an accuracy-oriented, interpretable ensemble for PE segmentation, suitable for offline or second-reader analysis.
• Faster single FCN backbones remain viable options for time-critical triage applications.
• The developed method shows promise for enhancing the detection of challenging pulmonary emboli.