InferVision
Mar 11, 2025
Sanford Health, the largest rural, not-for-profit healthcare system in the United States, is committed to bringing expert care closer to the 2.74 million people it serves in the upper Midwest. With 56 hospitals and nearly 300 clinics, Sanford Health’s network of more than 4,500 physicians and advanced practice providers is dedicated to providing advanced, expanded services. This commitment to innovation extends to Sanford Health’s research, where it explores cutting-edge technologies like artificial intelligence (AI) to improve diagnosis and patient care. A recent study led by Sanford Health researchers in Bismarck, ND, and presented at The Royal College of Radiologists Global AI Conference in London, England, demonstrated the potential for AI to revolutionize the detection of cancerous lymph nodes, a critical step in cancer staging.
The study focused on the development and evaluation of an AI-powered algorithm designed to detect and segment pathological lymph nodes in chest CT scans. Lymph nodes are small, bean-shaped organs that play an important role in the body’s immune system. However, they can also be a site of cancer spread, making their accurate identification vital for effective treatment planning. Mediastinal lymph nodes, located between the lungs in the chest, are particularly difficult to visualize on CT scans due to low tissue contrast and inherent variations in size, shape, and location. These challenges can lead to missed diagnoses or delayed treatment.
To solve this problem, the Sanford Health research team used an AI system developed by their partner Infervision that combines a RetinaNet detection algorithm (using ResNet50 as its backbone) with a UNet segmentation model. The AI was trained on a large dataset of 1,600 chest CT scans, allowing it to learn the subtle patterns and characteristics of both healthy and cancerous lymph nodes. This extensive training data is crucial to the AI’s accuracy and reliability.
The AI’s performance was then rigorously tested using a separate set of 211 CT scans that were carefully annotated by board-certified radiologists. This retrospective dataset provided a benchmark against which the AI’s findings could be compared. Instead of focusing on overall patient outcomes, the study used “station-by-station recall,” meaning that the AI’s ability to detect abnormalities in specific lymph node regions was evaluated. This approach provides a more clinically relevant assessment of the AI’s performance because it reflects how radiologists analyze CT scans in practice.
The results of the study were quite encouraging. The AI system achieved a promising station-specific sensitivity of 93.1%, meaning that it correctly identified 93.1% of cancerous lymph nodes in the specified regions. In practical terms, this meant that the AI correctly identified 378 true positive cases, missing only 28. Although the AI produced 46 false positives, identifying benign nodes or other structures as cancerous, this rate is relatively low and represents a significant improvement over previous algorithms. More importantly, the AI model maintained or even exceeded the sensitivity of previous algorithms, which ranged from 52.9% to 95.5%, while simultaneously reducing the number of false positives.
This improvement is likely attributable to the larger training dataset used in this study. The implications of these findings are significant. This AI-powered tool has the potential to significantly improve the accuracy and efficiency of cancer staging. By helping radiologists identify cancerous lymph nodes, AI could contribute to earlier diagnoses, more personalized treatment plans, and ultimately better patient outcomes. The station-specific approach used in the study provides a more clinically relevant assessment of AI’s performance, further strengthening its potential for real-world application.
While these results are promising, the researchers emphasize that larger-scale studies are needed to further develop the technology and validate its broader clinical applications.
Integrating AI into the clinical workflow requires careful planning and testing to ensure its seamless and effective implementation. However, the study conducted by Sanford Health researchers represents a significant step forward in the fight against cancer and demonstrates the power of AI to enhance diagnostic capabilities and improve patient care. As technology continues to advance, AI-driven solutions like this have tremendous potential to transform healthcare and improve the lives of countless individuals.