Objective: The meniscus tear is a common problem in sports trauma. The imaging diagnosis mainly depends on the MRI. To improve the diagnostic accuracy and efficiency, a deep learning model was employed in this study and the identification efficiency has been evaluated.
Methods: The standard knee MRI images of 924 individual patients were used to complete the training, validation, and testing process. The Mask R-CNN was considered as the deep learning network structure, and the ResNet50 was considered as the backbone network. The deep learning model was trained and validated with a dataset containing 504 and 220 patients, respectively. The accuracy testing was performed on a dataset of 200 patients and reviewed by an experienced radiologist and a sports medicine physician.
Results: After training and validation, the deep learning model effectively recognized the healthy and injured meniscus. The overall average precision of the bounding box and pixel mask was more than 88% when the IoU threshold value was 0.75. The detailed average precision of three types of menisci (healthy, torn, and degenerated) was ranged from 68% to 80%. The overall sensitivity of the bounding box and pixel mask was more than 74% at the IoU threshold from 0.50 to 0.95. The diagnosis accuracy for the healthy, torn, and degenerated meniscus was 87.50%, 86.96%, and 84.78%, respectively.
Conclusion: The Mask R-CNN recognized effectively and predicted the meniscus injury, especially for the tears that occurred at different parts of the meniscus. The recognition accuracy was admirable. The diagnostic accuracy can be further improved with the increase of the training sample size. Therefore, this tool has great potential in the application for the diagnosis of meniscus injury.