Abstract
Deep Vein Thrombosis (DVT) is a blood clot most commonly found in the leg, which can lead to fatal pulmonary embolism (PE). Compression ultrasound (CUS) of the legs is the diagnostic gold standard, but the high volume of negative referrals creates a significant clinical burden and long waiting times. We present a deep learning approach for the automatic interpretation of CUS images, designed to guide non-specialists at the point of care. Trained on 246 healthy volunteers and evaluated on 51 prospectively enrolled patients, our algorithm achieved a sensitivity of 93.8% and a specificity of 84.2%. Our analysis suggests the system is cost-effective at up to $150 per software examination, potentially transforming DVT diagnostic pathways by moving screening to the front lines of primary care.
Introduction
Deep Vein Thrombosis (DVT) represents a major global health challenge. When a blood clot forms in the deep veins of the leg, it poses a risk of dislodging and causing a pulmonary embolism—a potentially fatal condition. Currently, the "gold standard" for diagnosis is compression ultrasound (CUS). In this procedure, a specialist applies pressure with an ultrasound probe; a healthy vein collapses under pressure, while a vein containing a clot remains distended.
Despite its effectiveness, CUS is limited by its dependence on highly trained radiologists or sonographers. This dependency creates a bottleneck in healthcare systems: approximately 85–90% of patients referred for suspected DVT are ultimately found to be negative. These referrals lead to long wait times and increased clinical load. There is a clear need for a point-of-care solution that allows non-specialists, such as general practitioners or nurses, to reliably screen for DVT and triage patients effectively.
The AutoDVT System
To address this, we developed AutoDVT, an AI-powered software system designed to guide users through the CUS process. Unlike traditional image recognition models, AutoDVT acts as a real-time assistant, performing three integrated tasks simultaneously:
- Vessel Masking: Utilizing a U-Net backbone, the system segments veins and arteries within the ultrasound feed.
- Anatomical Landmarking: An auxiliary branch identifies the probe’s location across 11 key anatomical landmarks (e.g., the common femoral or popliteal veins), ensuring the examination covers the necessary vascular regions.
- Compression Analysis: A dedicated branch evaluates whether the vein is "open" or "fully closed" during the physical act of compression.
A critical innovation in AutoDVT is its use of spatio-temporal consistency. Rather than analyzing isolated frames, the model processes a sliding window of nine consecutive frames. This approach prevents "flickering" in the AI's output and maintains a stable interpretation of the vessel's state as the user applies and releases pressure.
Experimental Methodology
The development and validation of AutoDVT followed a multi-stage process to ensure robustness across different clinical environments and hardware.
Data Collection and Training
The model was trained on a substantial dataset of 167,145 annotated frames derived from 246 healthy volunteers and 9 DVT patients. Internal validation was conducted using an additional 15,523 frames. To ensure the AI was not overfitted to specific hardware, the team utilized various mobile ultrasound devices, including the Clarius L7, Philips Lumify, and GE VScan.
Clinical Evaluation
The system's performance was evaluated on a prospective cohort of 51 patients at an NHS DVT diagnostic clinic. This group included 32 DVT-positive and 19 DVT-negative patients. The AI's findings were compared directly against the clinical gold standard—an ultrasound performed by a specialist.
Key Results
AutoDVT demonstrated high diagnostic accuracy and efficiency, making it a viable candidate for clinical triaging:
- Sensitivity: 93.8% (Ability to correctly identify DVT)
- Specificity: 84.2% (Ability to correctly identify healthy veins)
- Negative Predictive Value (NPV): 88.9%
- Processing Speed: 25ms per frame, enabling real-time feedback.
The high NPV is particularly significant for primary care. It suggests that the tool can safely "rule out" DVT in a large portion of the population, preventing unnecessary anticoagulant injections and reducing the volume of referrals to specialist departments.
Discussion and Economic Impact
The "Tutor" Paradigm
One of the most important insights from this research is the shift from AI as a "judge" to AI as a "tutor." By providing real-time feedback on probe placement and compression technique, AutoDVT standardizes the examination style. This guidance compensates for the user's lack of specialized training, effectively lowering the barrier to entry for performing reliable ultrasound scans.
Economic Viability
The study included a cost-effectiveness analysis, evaluating the entire clinical DVT decision algorithm. The results indicate that the system remains economically viable at a cost of up to $150 per examination, assuming a standard willingness to pay of $26,000 per quality-adjusted life year (QALY). This suggests that integrating AutoDVT into the standard diagnostic pathway could provide both clinical and financial benefits to healthcare providers.
Limitations and Future Work
While the results are promising, the study has limitations. The external validation cohort (n=51) is relatively small, and a larger multi-centre prospective trial is currently underway to confirm these findings. Additionally, while the AI provides guidance, ultrasound remains a free-hand technique; operator variability still plays a role in the quality of the data captured. Future work will focus on a fully "hands-off" prospective study where non-specialists use the tool independently from start to finish.
Conclusion
AutoDVT represents a significant step toward democratizing specialized medical diagnostics. By combining deep learning with real-time user guidance, the system enables non-specialists to perform high-quality DVT screenings, potentially alleviating the burden on radiology departments and improving patient outcomes through faster point-of-care diagnosis.
Summary updated