Deep Learning-Based Image Segmentation In The Treatment Of Anterior Cruciate Ligament Injury

Anterior Cruciate Ligament(ACL)injury is one of the common knee injuries,which can seriously affect the knee stability and joint function of patients.At present,Anterior Cruciate Ligament Reconstruction(ACLR)has been widely used in the clinical treatment of ACL injury.Doctors need to evaluate the patient’s anterior cruciate ligament injury before surgery to achieve better surgical results.Arthroscopy,as a common method for diagnosing ACL injury,has high sensitivity,but it is an invasive examination.In recent years,Magnetic Resonance Imaging(MRI)has become a very effective and non-invasive method for clinical diagnosis of ACL injury.However,in clinical practice,there are problems such as the need for manual segmentation of the region of interest by doctors,the accuracy of image segmentation depends on the experience of doctors,and the high cost of segmentation time.Therefore,this thesis will study a deep learning-based method for ROI segmentation of knee MR images to achieve automatic segmentation of MR images,and on this basis,discuss the potential value of deep learning for the diagnosis and treatment of ACL injuries.The research work of this thesis is as follows:(1)Because the volume of the intercondylar notch structure in the knee joint is highly correlated with ACL injury,automatic segmentation of the intercondylar notch is of great significance for the prediction and prevention of ACL injury and the design of surgical plans.Because the intercondylar fossa has a complex three-dimensional anatomical structure and the shape of the two-dimensional section is obviously different,if the segmentation effect is not good,it will cause a great deviation in the final volume result.For this reason,we use the Res-UNet network to establish an automatic segmentation system of the intercondylar fossa based on MR images.The system can automatically segment the intercondylar notch structure and measure the intercondylar notch volume from knee MR images.From the experimental results,the segmentation index DSC reaches 0.916,and the relative error of the volume measurement is only 4.7%,indicating that the system has good segmentation performance and meets the accuracy required by clinical practice.And through clinical statistical analysis,the correlation between smaller intercondylar fossa volume and ACL injury was determined,which provided technical support for clinical prediction and prevention of ACL injury.(2)Preoperative prediction of the diameter of the graft used in ACLR has important clinical significance for guiding the design of ACLR and improving the surgical outcome.The hamstring tendon is a commonly used graft,and studies have shown that hamstring tendon cross-sectional area measured by MRI can predict intraoperative graft diameter.Due to the very small cross-sectional area of the hamstring tendon,slight deviations in the segmentation results will cause large errors in the area measurement.To this end,this thesis proposes an automatic segmentation model for hamstring tendon MRI based on U-Net++network with edge guidance module.From the experimental results,the segmentation index DSC reaches 0.922,and the relative error of tendon cross-sectional area measurement is only 8.3%,indicating that the performance of the system can meet clinical needs.In addition,the correlation between graft diameter size and hamstring cross-sectional area was validated by clinical analysis.The model can quickly and accurately segment tendon structures from MRI images and measure the cross-sectional area of the semitendinosus and gracilis tendon before surgery,guiding surgeons in better preoperative planning to determine the best graft choose.