Anterior cruciate ligament injury intelligent detection via slice standardization and local feature attention

Anterior cruciate ligament (ACL) injury can lead to loss of joint stability and flexion, seriously affecting knee motor function. Magnetic resonance imaging (MRI) is the preferred technique for evaluating ACL injury. However, interpreting an MRI of the ACL requires a high degree of musculoskeletal radiology experience for the clinicians due to the complexity and small anatomical structure of ACL, resulting in low consistency of inter-observer agreement and low accuracy of injury classification.

To improve the consistency of inter-observer agreement and accuracy of injury classification, this paper proposes a novel multi-instance learning framework based on the slice standardization and local feature attention.

To begin with, we construct the slice standardization module (SSM) by combining two feature integration methods: superimposing and normalizing the features of single-plane MRI slices, and logically connecting the features from three planes. Then, we design the Local Feature Attention Module (LFAM) by integrating a Channel Attention Module, a Spatial Attention Module, and residual connections. Finally, we develop a classification network for ACL injuries, named ACLNet, by combining SSM, LFAM, three fully connected layers, and a sigmoid layer. Experiments are conducted based on a single-center dataset. The dataset includes the MRI examination data of the knee joints from 264 cases. Each case includes MR images from three planes: axial, sagittal, and coronal. The training dataset includes 184 cases. The testing dataset consists of 80 cases. The MRI scans are performed using 1.5T or 3.0T scanners. The images are obtained with specific contrast weighting: axial and sagittal images are proton density weighted fat-suppressed imaging, while coronal images are T2-weighted fat-suppressed imaging.

On the testing dataset, ACLNet achieves 0.962, 0.987, 0.937, 1.000 and 0.967, in terms of accuracy, AUC, recall rate, precision, and F1 score, respectively. By replacing AlexNet in ACLNet with several classic deep classification networks, the choice of AlexNet is found to be preferable. Comparative analysis with several classic deep classification networks, transformer-based models, and knee injury classification models demonstrates that ACLNet can extract ACL injury classification information comparable to clinicians and perform best in automatically detecting ACL injuries on the dataset collected in this study.

Our method performs well on anterior cruciate ligament injury detection, which may have potential to make diagnosis quicker and more accurate.