World Thrombosis Day 2026: Progress in Artificial Intelligence for Point of Care Deep Vein Thrombosis Detection

Deep vein thromboses (DVT) can have serious consequences, ranging from pulmonary embolism (PE), to debilitating post-thrombotic syndrome. DVT and PE together affect roughly 10 million people globally every year.¹ Our field of vascular surgery frequently encounters DVT along several stages of the clinical pathway, from interpreting the ultrasound studies that diagnose DVT, to performing interventions for venous disease. This often gives us a clear view of bottlenecks that impede this pathway, most notably in diagnosis. Clinicians typically rely on venous duplex ultrasound to diagnose DVT, but access to imaging specialists performing these scans can be a challenge around the globe. This is particularly true during nights and weekends, or in under-resourced areas. Three quarters of American hospitals lack ultrasound coverage overnight, and only a third of hospitals in rural areas of the United States offer ultrasound services at all.^2,3Across the United Kingdom, patients can wait up to 48 hours for venous ultrasounds (the UK National Health Service target wait time is less than 4 hours).^4,5  Global shortages of sonographers and ultrasound equipment also hamper care delivery.^6,7

To explore ways to mitigate this imaging bottleneck , we examined an artificial intelligence (AI) ultrasound guidance system (ThinkSono Guidance, ThinkSono GmbH, Potsdam, Germany). The studied system, developed through collaboration between engineers, researchers, and clinicians from around the world, is a convolutional neural network which provides semantic segmentation masks for arteries and veins on ultrasound images. It is trained to identify blood vessels, anatomic landmarks, and vessel compression states. This system allows medical staff without formal ultrasound training to perform point of care venous compression ultrasound scans for lower extremity DVT.

The AI system does not make any diagnoses. It provides real-time feedback on probe position and compression technique to guide the operator to acquire ultrasound images, and then transmits these images to qualified clinicians for interpretation. Patients can then be triaged appropriately; any patients assessed to have incompressible veins (i.e. likely to have DVT) are then sent for confirmatory ultrasound by an imaging specialist (e.g. sonographer). The system is not designed to replace venous compression ultrasound by imaging specialists, but rather to function as a triage system that can extend the reach of these limited resources and improve access to and the efficiency of ultrasound assessment for lower extremity DVT.

Though this system had previously been studied in pre-clinical and controlled clinical settings, our study was one of the first to assess the feasibility of AI-guided ultrasound for DVT detection in a real-world setting, without a control specialist-performed venous compression ultrasound for every patient. Patients presenting to a hospital emergency department with suspected DVT were triaged using AI-guided scans reviewed by clinicians in conjunction with d-dimer testing. AI-guided scans were performed by non-ultrasound trained staff and ultrasound images were remotely reviewed by the on-call radiologist. All patients assessed as “compressible” (i.e. no DVT) with negative d-dimers were discharged. Patients with suspected DVT or indeterminate AI-guided scans were sent for a confirmatory venous compression ultrasound scan by an imaging specialist.

We found that across 53 patients, all AI-guided scans were of diagnostic image quality as rated by the interpreting clinician. Seventeen patients were able to be discharged and 28 were referred for specialist-performed ultrasound for elevated or unavailable d-dimer testing.  Eight patients were escalated to specialist-performed ultrasound for suspected DVT, of which 6 were confirmed to have DVT. The median time from AI-guided scan initiation to review was 38 minutes. AI-guided triage pathway sensitivity was 100%, and specificity was 96%. This provided proof-of-concept the AI guidance system could safely and rapidly triage suspected DVT patients, and that implementation in real-world settings was feasible, though it is worth noting the study was limited by a relatively small sample size.

Our study concluded recruitment roughly four years ago, and the studied system has evolved significantly since then. After receiving appropriate regulatory clearances, it is now in use across Europe, the UK, and the US. Initial results show promise in addressing challenges in DVT assessment. In an observational study of one urgent care site in the UK, for example, 120 patients were scanned using the AI system over three months. The percentage of patients receiving a scan within the recommended 4-hour window increased from 18% at baseline to 51% with AI-guidance system implementation. Average time from presentation to diagnosis was reduced from 20 hours to 9 hours. The number of required sonographer-performed ultrasounds was reduced by 80%. The median time from beginning an AI-guided scan to final result was 29 minutes.⁸ Though this again is only one relatively small study, it adds another encouraging data point in the advancement of this technology.

Though we are excited by this progress, much work remains in integrating this technology into clinical workflows for DVT detection. Efforts are ongoing to determine optimal implementation pathways tailored to the needs of different settings, from inpatient wards and emergency departments, to urgent care centers and paramedic use. Current development is also focusing on expanding clinical utility by adding features like doppler and guidance for calf vein scanning, which the version of the system in our study did not include. Through the continued collaboration of a broad and global group of colleagues, we hope to utilize technologies like this one to ensure every patient with a suspected DVT receives the care they need, regardless of time or clinical setting.

Disclosures:

This post: G. Speranza: ThinkSono Ltd., consultant, option holder

Original publication: E. Avgerinos: ThinkSono, Ltd., consultant, shareholder; G. Speranza: ThinkSono Ltd., consultant, option holder; F. Al Noor: ThinkSono, Ltd., employee, shareholder; S. Mischkewitz: ThinkSono, Ltd., employee, shareholder.

References

1.  Khan F, Tritschler T, Kahn SR, Rodger MA. Venous thromboembolism. Lancet. 2021 Jul 3;398(10294):64-77. doi: 10.1016/S0140-6736(20)32658-1. Epub 2021 May 10. PMID: 33984268.
2. Desser, T. S., Rubin, D. L. & Schraedley-Desmond, P. Coverage of emergency after-hours ultrasound cases: survey of practices at U.S. Teaching hospitals. Acad. Radiol.13, 249–253 (2006).
3. Khaliq, A. A., Nsiah, E., Bilal, N. H., Hughes, D. R. & Duszak, R. Jr. The scope and distribution of imaging services at critical access hospitals. J. Am. Coll. Radiol. 11, 857–862 (2014).
4. 4. NHS England. Venous compression ultrasound volumes and waiting times [Freedom of Information response]. London: NHS England; 2026.
5. https://www.nice.org.uk/guidance/ng158
6. Ginsburg, A.S., Liddy, Z., Khazaneh, P.T. et al. A survey of barriers and facilitators to ultrasound use in low- and middle-income countries. Sci Rep 13, 3322 (2023). https://doi.org/10.1038/s41598-023-30454-w
7. Naidoo G, Salim M, Jackson A, Handa A, Lakhoo K, Lindert J. Global survey on point-of-care ultrasound (pocus) use in child surgery. Pediatr Surg Int. 2024 Sep 5;40(1):249. doi: 10.1007/s00383-024-05797-8. PMID: 39237661; PMCID: PMC11377359.
8. Downing, J., Dattoo, M., Naqvi, D., West, H. Service Organisation & Design Evaluation of ThinkSono AI-Guided DVT Ultrasound Scanning to improve the SDEC DVT Pathway. Abstract accepted for presentation at: The Society for Acute Medicine, SAMGlasgow; Glasgow, Scotland; Oct 8-9, 2026.