Scientists from the Indian Institute of Science (IISc) Bengaluru, in partnership with a hospital, have created an artificial intelligence (AI) tool capable of recognising the median nerve in ultrasound videos and diagnosing carpal tunnel syndrome (CTS).
CTS occurs when the median nerve, travelling from the forearm to the hand, is compressed at the carpal tunnel part of the wrist. This compression leads to numbness, tingling, or pain. CTS is a prevalent nerve-related condition, particularly impacting individuals who perform repetitive hand movements, such as office workers using keyboards, assembly line employees, and athletes.
Presently, physicians use ultrasound for visualising the median nerve and assessing factors such as its size, shape, and potential irregularities. However, distinguishing details in ultrasound images and videos is challenging compared to X-rays and MRI scans.
According to a researcher, at the wrist, the nerve is quite visible, and its boundaries are clear. However, when examining the elbow region, numerous other structures are present, making the nerve boundaries less distinct. The accurate tracking of the median nerve is crucial, especially for treatments involving local anesthesia administration to the forearm or blocking the median nerve to alleviate pain.
In creating their tool, the team used a machine learning model using a transformer architecture, akin to the technology driving ChatGPT. Initially designed to simultaneously identify numerous objects on an online video-sharing and social media platform, the team streamlined the model by eliminating computationally intensive components and narrowing its focus to tracking a single object—the median nerve in this instance.
Working with the lead consultant neurologist at the hospital, the team collected and annotated ultrasound videos from both healthy individuals and those with CTS to train the model. Once trained, the model successfully segmented the median nerve into individual frames of the ultrasound video.
Additionally, the model demonstrated the capacity to automatically calculate the cross-sectional area of the nerve, a crucial factor in diagnosing CTS. Traditionally, this measurement is carried out manually by a sonographer. The tool automates this process and can measure the cross-sectional area in real-time. The tool was able to report the cross-sectional area of the median nerve with more than 95% accuracy at the wrist region.
While numerous machine learning models have been created for screening CT and MRI scans, there is a scarcity of models specifically designed for ultrasound videos, particularly in the context of nerve ultrasound.
The team initially trained the model on a single nerve, and now it plans to expand its capabilities to encompass all nerves in both the upper and lower limbs. Furthermore, the model has undergone a pilot test in the hospital. The hospital has an ultrasound machine linked to an extra monitor where the model is operational. While doctors observe the nerve, the software tool simultaneously outlines the nerve. Professionals can assess its real-time performance.
In the next phase, the team will identify ultrasound machine manufacturers interested in incorporating this tool into their systems. This type of tool can aid doctors and decrease the inference time. However, it’s essential to note that the ultimate diagnosis still requires the expertise of a physician. The findings have been published in IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control.