AI system able to identify standard plane, segment tissue in ultrasounds of shoulder
Researchers out of China have developed an AI system for shoulder ultrasound imaging, assisting sonographers looking for injuries in the rotator cuff. The deep-learning application was able to automatically identify the standard plane and segment tissue for improved diagnosis of injury. The results are published in Ultrasound in Medicine & Biology. [1]
“The identification and segmentation of structures in rotator cuff ultrasound scanning pose a significant challenge for novice practitioners. Although existing musculoskeletal ultrasound image segmentation research has focused primarily on isolating individual muscles or nerves, this study was aimed at identifying and segmenting multiple tissues within a single plane,” the study authors led by Rui Tang, PhD, of Peking University Third Hospital in Beijing wrote.
To measure how well the AI would perform identifying damage to the rotator cuff using exclusively ultrasound, the research team designed a standard plane recognition module for an existing image recognition model, ResNet50. For the automatic tissue segmentation piece of the study, a module was built using the Mask R-CNN model, an AI application designed for image segmentation.
Each new module was trained on a dataset of shoulder ultrasound images. While most deep-learning models can find the standard plane or segment tissue in a scan, few do both. Tang said the combined-AI system used for this research “demonstrated superior performance compared with similar studies for both functions.” He and the other authors added that, typically, those studies have used a different imaging modality other than ultrasound to detect rotator cuff tears.
The standard plane recognition module achieved an impressive 94.9% recognition accuracy in the test set, accompanied by an average precision rate of 96.4%, a recall rate of 95.4%, and an F1 score of 95.9% using a dataset of 59,265 shoulder joint ultrasound images. Meanwhile, the automatic tissue segmentation module, trained on 1886 images, exhibited an average intersection over union value of 96.2%.
Notably, the AI developed by Tang and his team consistently achieved intersection over union values exceeding 90.0% for all standard planes, indicating its effectiveness in precisely delineating anatomical structures.
“When compared to traditional approaches, the system demonstrates the ability to rapidly classify standard planes and automatically locate and segment tissues without requiring manual intervention,” the authors concluded. “This is expected to significantly reduce the learning curve associated with shoulder joint ultrasound, while simultaneously enhancing the quality and efficiency of sonographers in screening for shoulder joint diseases.”
You can read the full study here.
Chad is an award-winning writer and editor with over 15 years of experience working in media. He has a decade-long professional background in healthcare, working as a writer and in public relations.