Identification and diagnosis of meniscus tear by magnetic resonance imaging using a deep learning model

1. Introduction

Meniscus is commonly referred to as the fibrocartilaginous structure located within the knee joint cavity, between the femur and tibia, providing strength to the joint and absorbing impact for protection [1,2]. It can be divided into medial meniscus and lateral meniscus. Meniscus injury is very common, with an incidence rate of 6-7 in 10,000 [3]. Destruction of meniscal integrity due to various conditions such as dysplasia, chronic strain, and acute sprains can lead to meniscal damage, accompanied by a series of clinical symptoms such as pain and dysfunction that severely impact the patient’s mobility and quality of life. Once a meniscal injury is diagnosed, most of the cases need surgical treatment. Accurate and timely preoperative diagnosis is of great significance.

Magnetic resonance imaging (MRI) generates high imaging resolution of soft-tissue. This method allows a clear view of the shape and internal structure of the meniscus, and is the preferred examination for the diagnosis of meniscus injuries [4,5]. Fat-suppressed fast spin-echo proton density-weighted image (FS FSE PDWI), which produces homogeneous hypointense on MRI sequences, is most commonly used in the detection of meniscal injuries. A multi-center study showed that analyzing the risk and prognosis of meniscal injury had important clinical implications [6]. However, the accuracy of MRI diagnosis is limited due to the following reasons. Firstly, several irregularly shaped tissues are situated around the meniscus. Secondly, the abnormal signal of a meniscal tear is so small that it is not easy to be spotted on images. Thirdly, the amount of MRI data can be extremely huge (about 100 images per patient). Fourthly, the accuracy of diagnosis is influenced by the doctor’s diagnostic experience. Furthermore, other subjective factors may also affect the diagnostic results.

See also  What Is Image-Guided Radiation Therapy (IGRT)?

In recent years, the application of artificial intelligence (AI) in the field of medical imaging has become a research hotspot, and it is believed that AI has the potential to provide accurate diagnosis and treatment. Deep learning and other AI applications can effectively improve the efficiency of data processing and reduce human errors through repetitive learning to identify disease patterns [7,8]. Traditional machine learning algorithms mainly include neural network, k-nearest neighbor, support vector machine, naive Bayes classifier, and random decision forest. These algorithms rely on the shallow features of artificial intelligence. One advantage of deep learning is that there is no need to specify the features manually, and machine can learn by itself through dataset training, bringing a breakthrough in image processing.

Great progress has been made in the in-depth analysis of knee MRI images using AI, but it is far less used in other critical conditions such as tumor, nerve damage and pulmonary nodules. Compared to bone and cartilage, the study on meniscus is limited because image segmentation and post-processing are not feasible. Among the AI studies regarding meniscal tears, most studies only analyzed the sagittal plane, and a few studies analyzed the sagittal plane, coronal plane, and cross-section simultaneously [9]. The areas under the curve (AUCs) for these studies ranged from 0.847 to 0.910 [10], meaning that this technology should be improved to increase the diagnostic accuracy by MRI.

Slice thickness is an important parameter in meniscus MRI examination. In previous studies, the scanning layer thickness ranged from 0.7 ​mm to 3.0 ​mm [5,11], which made the data sources lacking homogeneity. This study aimed to utilize the most commonly used sequences and scanning layer thicknesses in clinical practice for model training, providing a wider application range and benefiting future multi-center studies. After obtaining the feature map of meniscus MRI images through convolutional neural networks, Mask R-CNN was used to perform classification, regression, and pixel-level mask diagnosis. To verify the recognition accuracy of the deep learning model, the results were evaluated by experienced doctors in conjunction with arthroscopic surgery. We anticipated that this technology could serve as an effective tool for clinical MRI-assisted diagnosis of meniscal injuries.