Deep vein thrombosis can be diagnosed accurately with automatic ultrasound image analysis and deep neural networks

We discuss a deep learning algorithm that is able to equip healthcare professionals with ultrasound image interpretation skills that are required for the accurate diagnosis of deep vein thrombosis.
Deep vein thrombosis can be diagnosed accurately with automatic ultrasound image analysis and deep neural networks

Background: Deep vein thrombosis is a dangerous widespread disease: venous thromboembolism (VTE) [1] is a global phenomenon associated with the development of blood clots in the leg veins - deep vein thrombosis (DVT). More people die annually from VTE than from breast cancer, prostate cancer, AIDS, and car accidents combined [1]. DVT is painful, causes swelling, and can trigger other, even fatal, complications. Annually, 544,000 Europeans die as a result of DVT [2].

To avoid misdiagnosis, patients are referred to hospital for imaging at the slightest suspicion of DVT. In up to 93% of all cases, the suspicion of DVT is not confirmed. This leads to enormous financial burdens for the health care system and increases health risks for patients who actually suffer from DVT due to unnecessary waiting times. Furthermore, complex clinical diagnostic imaging techniques are not available everywhere, resulting in additional risky waiting times.

It is notoriously difficult to diagnose a DVT by clinical judgment alone. The standard approach to making a diagnosis of proximal DVT currently involves a diagnostic pathway combining pre-test probability, D-dimer (blood) testing, and compression ultrasonography (typically a two-point compression examination). 

Handheld ultrasound probes have recently become available. These probes have enabled ‘app-based’ compression ultrasonography to be performed without the need for bulky cart or laptop-based ultrasound machines. These new machines have a small form factor, meaning only the ultrasound probe is required for diagnostic purposes in conjunction with a smartphone or table.

At present, although the new handheld probes are smaller and are better suited for point of care diagnosis, they still require an experienced radiologist or sonographer to perform the two-point compression exam. This means that these devices can only be used wherever radiographers/radiologists are based most often i.e. hospital X-ray departments. In our study we present and test a software-based approach for guided ultrasound image acquisition, analysis and interpretation support. Our approach can be used together with handheld probes which has the potential to assist non-radiology specialist healthcare professionals (e.g. nurses, non-radiologist physicians, general practitioners and other allied healthcare professionals) to carry out the compression ultrasound exam with minimal training. 

Our Solution: Our manuscript, entitled “Non-invasive Diagnosis of Deep Vein Thrombosis from Ultrasound Imaging with Machine Learning”, describes our deep learning method, a clinical validation and  a health-economic analysis of deep-learning for diagnosing DVT with compression ultrasound in two different European clinics.  

The deep neural network’s task is to annotate vessels, find anatomical landmarks, and analyse vessel compression state automatically. DVT diagnosis is done by automatization of the standard clinical ultrasound compression algorithm in a heuristic computer programme, based on the biometrics acquired from the deep neural network during the scan. This deep neural network has been trained mainly on data from healthy volunteers (n=246, age range 18-84, BMI < 30) and compression sequences from consented patients with confirmed DVT (n=9). Being able to learn from predominantly healthy volunteers is important for population-wide low-prevalence diseases like DVT. Traditional (deep) machine learning would require approximately half of the training samples to be from the disease class, which would require continuous ultrasound video data from thousands of patients to reach accuracies that would be comparable to our method. 

Our Major Findings: Algorithmic DVT diagnosis performance results in a sensitivity within a 95% CI range of (0.82, 0.94), specificity of (0.70, 0.82), positive predictive value of (0.65, 0.89), and a negative predictive value of (0.99, 1.00)  when compared to the clinical gold standard.  For comparison, ultrasound has a sensitivity of 94% and a specificity of 97% for DVT detection when performed by specialised radiologists [3,4]. Two earlier studies reported sensitivity of 84.4 – 90.0% and specificity of 97.0 – 97.1% when intensely trained nurses and GPs were the ultrasound operators [5,6]. 

Our approach is estimated to generate a positive net monetary benefit at costs up to £71 to £139 per software-supported examination, assuming a willingness to pay of £20 000/QALY.

A positive test with our method will always lead to a confirmatory scan with an expert. Within this group, the expert’s chance of seeing an actual DVT positive patient is then more than 80%, which is notably higher than the prevalence in the general population. Increasing the pre-test probability for DVT will likely reciprocally increase the diagnostic utility and discriminatory power of the expert examination as well.

Literature and our own experiments show strong evidence that a DVT examination in primary care performed by non-experts is feasible. We would expect that rapid point of care diagnostics and wide availability of testing, which is conceivably enabled by our approach, would lead to timely treatment, decreased stress, and increased patient satisfaction.

Stage 4 AutoDVT prototype
Fig. 1: The Stage 4 prototype (AutoDVT) used in this study.

The potential benefits of the discussed approach  include:

  • Rapid access to a diagnostic test
  • Point of care testing, reducing the need for patients to attend hospital radiology departments and hospital emergency departments
  • Reduced need for prophylactic anticoagulation whilst waiting for a scan, thereby reducing the potential risk to patient of bleeding whilst taking an anticoagulant medication
  • Reduced cost for diagnosis to health care providers due to no longer requiring radiographer led services and high technology USS devices and reduced use of anticoagulation
  • Higher patient satisfaction with their diagnostic clinical pathway

We have developed this paper's approach to the point of a stage 4 prototype called AutoDVT (Fig. 1), which can be tested by contacting . Our study describes the first step of a larger clinical trial programme which we will use to ultimately evaluate the clinical efficacy of the AutoDVT software for diagnosis of proximal DVT. 


[1]    A. T. Cohen et al., “Venous thromboembolism (VTE) in Europe. The number of VTE events and associated morbidity and mortality,” Thromb. Haemost., vol. 98, no. 4, pp. 756–764, Oct. 2007.
[2]    J. A. Heit, “Poster 68,” presented at the American Society of Hematology, 47th Annual Meeting, Atlanta, GA, December 10-13, 2005.
[3] Goodacre S, Sampson F, Thomas S, van Beek E, Sutton A, Systematic review and meta-analysis of the diagnostic accuracy of ultrasonography for deep vein thrombosis, BMC medical imaging, Dec 1;5(1):6, (2005)
[4]    Zierler BK, Ultrasonography and diagnosis of venous thromboembolism. Circulation. Mar 30;109(12_suppl_1):I-9, (2004)
[5] Mumoli N, Vitale J, Cocciolo M, Cei M, Brondi B, et al., Accuracy of nurse‐performed compression ultrasonography in the diagnosis of proximal symptomatic deep vein thrombosis: a prospective cohort study. Journal of Thrombosis and Haemostasis, Apr;12(4):430-5, (2014)
[6]    Mumoli N, Vitale J, Giorgi-Pierfranceschi M, Sabatini S, Tulino R, et al., General practitioner–performed compression ultrasonography for diagnosis of deep vein thrombosis of the leg: a multicenter, prospective cohort study, The Annals of Family Medicine, Nov 1;15(6):535-9, (2017)

Please sign in or register for FREE

If you are a registered user on Nature Portfolio Health Community, please sign in