Multi-sequence knee magnetic resonance imaging (MRI) is a sophisticated non-invasive technique employed for diagnosing knee conditions. Despite its advantages, interpreting MRI results can be very labor-intensive and requires specialized expertise. A recent research effort has unveiled a new deep learning model aimed at classifying 12 prevalent knee abnormalities, thus improving both the efficiency and precision of diagnosis.
Multi-sequence knee magnetic resonance imaging (MRI) is a sophisticated non-invasive technique employed for diagnosing knee conditions. Despite its advantages, interpreting MRI results can be very labor-intensive and requires specialized expertise. A recent research effort from the School of Engineering at the Hong Kong University of Science and Technology (HKUST) has unveiled a new deep learning model aimed at classifying 12 prevalent knee abnormalities, thus improving both the efficiency and precision of diagnosis.
This study represents a collaboration between HKUST’s Smart Lab in Hong Kong and the Third Affiliated Hospital of Southern Medical University located in Guangzhou, China. Their findings were published recently in Nature Communications under the title “Learning Co-Plane Attention Across MRI Sequences for Diagnosing Twelve Types of Knee Abnormalities.”
The knee joint, being a complex hinge joint, serves as one of the principal load-bearing joints in the human body, fundamental for a variety of movements during everyday activities. Knee abnormalities can develop due to aging or injury, leading to pain and functional limitations. Therefore, it is essential to diagnose these abnormalities accurately to tailor treatment options and enhance the quality of life for patients.
The knee’s intricate anatomical structure means that variations in scanning parameters can yield differing outcomes. Moreover, certain subtle lesions may be easily missed by radiologists who may lack sufficient experience.
To tackle these issues, the research team, headed by Assistant Professor CHEN Hao from both the Department of Computer Science and Engineering and the Department of Chemical and Biological Engineering at HKUST, worked alongside five hospitals to gather data from 1,748 patients. This included MRI sequences such as T1-weighted (T1W), T2-weighted (T2W), and proton density-weighted (PDW) across sagittal, coronal, and axial planes. By integrating data obtained from arthroscopy—widely regarded as the definitive method for diagnosing knee issues—researchers performed a thorough analysis and pinpointed 12 prevalent types of knee abnormalities among these patients.
The team developed a deep learning model that employs Co-Plane Attention across MRI Sequences (CoPAS) to categorize these abnormalities. This model proficiently captured intensity differences across various scanning parameters and discerned intricate relationships with abnormality types by separating spatial features from each MRI sequence, resulting in high classification accuracy.
To evaluate the model’s effectiveness, simulated clinical testing was conducted, where radiologists were first asked to make independent diagnoses based solely on MRI scans. After a break, they were then requested to re-evaluate their diagnoses using the model’s output as a guideline.
Comparing the outcomes showed that the model achieved an average diagnostic accuracy that surpassed that of junior radiologists and was on par with senior radiologists. Overall, all radiologists demonstrated significantly improved accuracy with the aid of the model.
An additional interpretability analysis contrasted clinical empirical data with the model’s output. The results indicated that the model’s decision-making process aligned consistently with clinical practices, suggesting that it had developed a set of rules similar to those used by human radiologists, enabling it to deliver more reliable results in clinical scenarios.
Prof. Chen remarked, “The innovative CoPAS model exhibits diagnostic performance akin to that of radiologists. It is especially useful in bridging the experience gap between less seasoned and senior doctors.”
“Our findings highlight the potential of artificial intelligence in healthcare, illustrating its ability to uncover and validate new clinical insights,” he added.
Prof. Chen is among the corresponding authors of the paper, alongside Prof. ZHAO Yinghua from the Third Affiliated Hospital of Southern Medical University. Co-first authors include QIU Zelin, a doctoral student in Computer Science and Engineering at HKUST, and Dr. XIE Zhuoyao from the Third Affiliated Hospital of Southern Medical University.
