Leveraging AI for Deep Venous Thrombosis Detection:Clinical Advancements and Challenges

By Campion Quinn, MD

Introduction

Artificial intelligence (AI) has emerged as a transformative tool in medical diagnostics, offering new possibilities for rapid, precise disease detection. One such area of focus is the diagnosis of deep venous thrombosis (DVT), a potentially life-threatening condition that, if untreated, can lead to severe complications like pulmonary embolism or chronic venous insufficiency. Traditional methods of DVT diagnosis, including ultrasonography and computed tomography angiography (CTA), have limitations, such as variability between observers and diagnostic delays due to the lack of immediate radiologist availability. This essay explores how AI models are being developed to overcome these challenges, focusing on a recent study integrating AI into iliofemoral DVT detection from CTA imaging.

The Scope of DVT Detection in Clinical Practice

DVT primarily affects the lower extremities, with iliofemoral DVT involving the large veins of the pelvis and upper leg. Accurate and timely diagnosis is essential for preventing complications, but the symptoms can be nonspecific, often requiring imaging for confirmation. The common modalities include ultrasonography, CTA, and magnetic resonance imaging (MRI). However, radiologists need help with significant challenges, including time-consuming image review and intra-observer variability, especially when evaluating complex or ambiguous cases. As a result, AI has been explored as a supplementary tool to enhance diagnostic accuracy and efficiency​.

AI in DVT Detection: The Role of Deep Learning Models

The study by Seo et al.[1] investigates the potential of convolutional neural networks (CNNs), specifically RetinaNet, to detect iliofemoral DVT using CTA images. RetinaNet, known for its time efficiency and high accuracy, is designed as a one-stage detection model with a specialized loss function to improve object detection. The research applies this AI model to identify DVT by leveraging adjacent slices of imaging data, mimicking how radiologists interpret multiple slices for diagnostic confirmation.

Data Preparation and Model Performance

The research team analyzed data from 380 patients, split evenly between those with and without iliofemoral DVT, using a variety of training and validation approaches. The AI algorithm was trained to extract features from individual slices and synthesized sets of three consecutive slices to simulate a real-world clinical diagnostic process. The team employed ResNet-based models to optimize detection performance. The AI models achieved high sensitivity and precision, with the highest sensitivity reaching 84.3% when using synthesized three-slice data with the ResNet50 backbone. This approach demonstrated that incorporating multiple images improved diagnostic performance by capturing more contextual information despite slightly increasing false positives​.

Clinical Advantages of AI-Assisted DVT Detection

The primary advantage of AI-enhanced detection is the potential to reduce radiologists' diagnostic workload while maintaining high accuracy. In scenarios where immediate radiologist review is impossible, such as during night shifts, AI algorithms can provide preliminary assessments, highlighting potential DVT cases for further review. The study showed that the AI model mimicked radiologists' diagnostic approaches and reduced the time required for review, potentially improving patient outcomes by accelerating diagnosis and treatment initiation​.

Moreover, AI can address the variability inherent in manual interpretation. The study’s AI model performed consistently across multiple datasets, reducing the risks associated with human error and fatigue. This consistency is particularly valuable in high-stress environments such as emergency departments, where rapid decision-making is critical.

Challenges and Limitations

Despite its promise, AI-assisted DVT detection is challenging. One significant limitation highlighted in the study is the potential for false positives, especially when the AI model encounters structures with similar shapes or densities as DVTs. False positives can increase the workload for radiologists, requiring them to evaluate flagged cases carefully. Additionally, the AI model's performance can vary based on the imaging modality, with synthesized data sets showing higher sensitivity but slightly lower precision than single-slice evaluations​.

Another challenge is the generalizability of AI algorithms across diverse patient populations. The dataset used in the study was drawn from a specific clinical setting, and further validation is necessary to determine whether the AI model performs equally well in other healthcare environments. Bias in data selection, such as excluding patients with artifacts or incomplete data, may also impact the algorithm’s effectiveness when applied in real-world settings.

Future Directions

The study suggests several avenues for further research. Expanding the detection range beyond iliofemoral veins to include infrapopliteal veins could enhance the clinical value of AI models, especially for high-risk patients. Additionally, the research emphasizes the need to compare AI-assisted diagnoses with those made by experienced radiologists to assess the practical advantages of AI integration in clinical workflows. The authors also advocate for incorporating more diverse datasets to improve the robustness and applicability of AI models across different patient populations and healthcare settings​.

Conclusion

Integrating AI into DVT detection represents a promising advancement in medical diagnostics. By leveraging CNN-based models like RetinaNet, AI can assist radiologists in making faster and more accurate diagnoses, particularly in time-sensitive scenarios such as emergency care. The study by Seo et al. demonstrates that AI has the potential to complement radiological expertise, improving diagnostic efficiency and reducing variability in clinical practice. However, challenges related to false positives, data biases, and the need for extensive validation remain. Addressing these issues through further research and technological refinement will be essential to unlocking the full potential of AI in DVT detection and other areas of medicine.

As healthcare systems increasingly adopt AI technologies, physicians and radiologists must stay informed about their capabilities and limitations. Understanding how AI algorithms work and collaborating with developers to refine their applications will ensure that AI is a reliable ally in improving patient outcomes. This study offers a glimpse into the future of diagnostic medicine, where AI and clinical expertise converge to provide faster, more accurate care.

If you’re interested in the medical applications of AI and want to learn more, consider subscribing to my free newsletter, AI in Medicine, by clicking here:  https://aiinmedicine.beehiiv.com/subscribe

[1] Seo, J.W., Park, S., Kim, Y.J. et al. Artificial intelligence-based iliofemoral deep venous thrombosis detection using a clinical approach. Sci Rep 13, 967 (2023). https://doi.org/10.1038/s41598-022-25849-0