HUCSR-Net: Automated Pulmonary Embolism Detection Using 3D Deep Learning

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Pulmonary embolism (PE) is a significant cause of cardiovascular mortality and a diagnostic challenge.
• Computed tomography pulmonary angiography (CTPA) is the standard but is time-consuming and requires expert interpretation.
• Automated diagnostic tools are needed to expedite treatment and improve outcomes.

Purpose of the Study:

• To develop and validate HUCSR-Net, an automated 3D convolutional neural network for PE detection.
• To assess the model's performance using imaging studies from a single center.
• To explore the feasibility of AI-driven PE detection in resource-constrained settings.

Main Methods:

• Developed HUCSR-Net, a 3D convolutional neural network based on R(2 + 1)D-18 architecture.
• Trained the model on 128,484 imaging studies from 86 patients using AdamW optimizer and binary cross-entropy loss.
• Employed 5-fold cross-validation and evaluated performance using AUC, sensitivity, specificity, and F1-score.

Main Results:

• The optimal model achieved a validation AUC of 0.7026, sensitivity of 0.8592, and F1-score of 0.592.
• Demonstrated solid discriminatory performance with a Matthews correlation coefficient of 0.1515.
• Results confirm the model's diagnostic capability despite a modest cohort size.

Conclusions:

• A deep 3D convolutional network trained on a single-center cohort can achieve diagnostic performance comparable to large-scale studies.
• HUCSR-Net shows feasibility for automated PE detection in resource-constrained environments.
• The study supports integrating such AI systems as decision-support tools for radiologists.