Deep Learning Approaches for Thrombosis Detection and Risk Assessment Via Ultrasound Imaging: A Scoping Review.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Medicine

Background:

• Thrombosis, or blood clot formation, presents significant health risks like pulmonary embolism.
• Ultrasound (US) imaging is a key diagnostic tool, but its effectiveness can be limited by operator dependence.
• Deep learning (DL) offers potential to overcome these limitations in US-based thrombosis assessment.

Purpose of the Study:

• To review the application of deep learning (DL) techniques in enhancing thrombosis detection and risk assessment using ultrasound (US) imaging.
• To explore DL's role across venous, arterial, and cardiac contexts for thrombosis diagnosis.
• To identify common DL models and their performance in US-based thrombosis studies.

Main Methods:

• A comprehensive scoping review of literature from PubMed and Scopus was performed.
• Studies utilizing DL models for thrombus detection, classification, segmentation, or risk prediction with vascular US were included.
• PRISMA-ScR methodology guided the literature search and selection process.

Main Results:

• Twenty-two studies met eligibility criteria, predominantly using Convolutional Neural Networks (CNNs), U-Net, ResNet, and Artificial Neural Networks (ANNs).
• DL models showed promise in deep vein thrombosis (DVT) diagnosis, arterial plaque segmentation, and cardiac thrombus differentiation.
• Reported high sensitivity, specificity, accuracy, and AUC, often surpassing traditional methods, despite dataset variability.

Conclusions:

• Deep learning-enhanced ultrasound imaging demonstrates significant potential for improving diagnostic precision in thrombosis care.
• DL applications show promise in venous, arterial, and cardiac thrombosis diagnostics, aiding clinical decision-making.
• Future research should focus on model interpretability, real-world integration, and standardized datasets.