Deep Learning Approaches for Thrombosis Detection and Risk Assessment via Ultrasound Imaging

Thrombosis, the formation of a blood clot inside a blood vessel, is a critical condition with significant health risks, including pulmonary embolism (PE) and chronic post-thrombotic syndrome. Deep learning (DL) approaches have shown promise in enhancing the detection and risk assessment of thrombosis, addressing key challenges in ultrasound (US) imaging interpretation. US imaging remains the method of choice for the initial assessment and monitoring of thrombosis due to its non-invasiveness, real-time capability, and absence of ionizing radiation. However, despite being a cost-effective and widely available tool, its diagnostic accuracy heavily depends on operator expertise, leading to variability in image acquisition and interpretation. The integration of DL techniques in US imaging has the potential to improve diagnostic precision, automate analysis, and reduce operator dependency. This thesis systematically reviews the role of DL in thrombosis detection and risk assessment via US imaging, categorizing approaches based on venous, arterial, and cardiac thrombosis.

A systematic search across PubMed and Scopus identified 233 studies, of which 22 met the inclusion criteria. The review highlights that convolutional neural networks (CNNs), U-Net, ResNet, and artificial neural networks (ANNs) are the most frequently employed models for classification, segmentation, and feature extraction. The study further categorizes DL-based thrombosis detection based on prediction tasks, including thrombus classification, vessel segmentation, flow analysis, and thrombus localization.

For venous thrombosis, DL models have been utilized to detect deep vein thrombosis (DVT) by evaluating vein compressibility, classifying thrombi, and assisting non-specialists in remote diagnosis. AI-guided freehand point-of-care ultrasound (POCUS) methods have been explored, demonstrating high sensitivity and specificity in detecting DVT.

For arterial thrombosis, CNN-based approaches have been applied to classify and segment atherosclerotic plaques, including the detection of vulnerable plaques associated with acute coronary syndrome (ACS). Deep learning models have also been integrated with intravascular ultrasound (IVUS) to enhance 3D vessel reconstruction, aiding in the assessment of plaque burden and stenosis severity. AI-based segmentation models, such as U-Net and Mask-RCNN, have significantly improved vascular lesion classification.

For cardiac thrombosis, DL techniques have been applied in transesophageal echocardiography (TEE) for the classification and segmentation of intracardiac masses, distinguishing thrombi from tumors. Computer-aided diagnostic (CAD) algorithms have enhanced thrombus detection in patients with atrial fibrillation, demonstrating improved sensitivity and specificity when combined with expert evaluation.

Overall, DL-based approaches for thrombosis detection and risk assessment using US imaging have shown considerable advancements in diagnostic accuracy, automation of image analysis, and clinical decision support. However, challenges remain, including dataset availability, variability in US image quality, and the need for multi-center validation. Future research should prioritize real-world clinical integration, model interpretability, and the development of standardized, publicly available datasets for thrombosis assessment. This systematic review provides insights into the current state, challenges, and future directions of AI-driven vascular imaging, highlighting its potential impact in thrombosis detection and risk assessment across venous, arterial, and cardiac applications.