- Applying a deep learning model to knee magnetic resonance imaging (MRI) exams helped increase the accuracy and speed of knee injury classification, and could provide clinical decision support for radiologists, a study published in PLOS Medicine revealed.
More MRI examinations are performed on the knee than on any other region of the body, the researchers said, resulting in a high volume of images that need to be assessed for abnormalities. In addition, the level of detail in these images can make accurate interpretation a time-consuming process.
To accelerate MRI interpretation and enhance diagnostic accuracy, the team developed a deep learning model using reports for knee MRI exams between 2001 and 2012, and set out to compare its performance to that of radiologists.
The results showed that the model was able to accurately classify pathologies on knee MRIs, achieving an area under the receiver curve (AUC) of 0.937 for abnormality detection.
Specifically, the model achieved an AUC of 0.965 for ACL tear detection, and an AUC of 0.847 for meniscus tear detection.
Researchers also found that the model reviewed images significantly faster than human radiologists. While providers required more than three hours on average to interpret 120 exams, the model delivered all classifications in under two minutes.
Researchers also wanted to see if radiologists’ diagnostic accuracy would improve when assisted by the deep learning model. The team displayed the model’s findings to radiologists and asked whether human clinicians agreed with them.
Researchers found that clinicians’ specificity for identifying ACL tears improved by 4.8 percent when shown the model’s findings. This increase in specificity could mean that fewer patients receive unnecessary surgery for potential ACL tears that haven’t been confirmed by MRI exams.
The findings show that deep learning models could help accelerate MRI interpretation, and could serve as clinical decision support tools for radiologists.
“Our results demonstrate that a deep learning approach can achieve high performance in clinical classification tasks on knee MRIs,” the researchers said. “We also found that providing the deep learning model predictions to human clinical experts as a diagnostic aid may improve clinical interpretations.”
This deep learning model has many potential applications in clinical practice, the researchers said, including optimizing radiologists’ workflows and improving care quality.
“Providing rapid results to the ordering clinician could improve the efficiency of the healthcare system,” the group stated.
“Automated abnormality prediction and localization could help general radiologists interpret medical imaging for patients at the point of care rather than waiting for specialized radiologist interpretation, which could aid in efficient interpretation and reduce errors.”
However, the knee injury study did have some limitations, the researchers noted. The deep learning model was trained on MRI data from one institution, and future research will need to determine whether models trained on larger and multi-institutional datasets can achieve similar results.
Additionally, the team said that future studies should include a larger group of radiologists to better evaluate the impact of the deep learning model on clinicians’ performance.
Still, the findings from this study show that deep learning approaches are viable tools for improving MRI interpretation and clinical diagnosis.
“Our results provide early support for a future where deep learning models may play a significant role in assisting clinicians and healthcare systems,” the team concluded. “More studies are necessary to evaluate the optimal integration of this model and other deep learning models in the clinical setting.”