Musculoskeletal Tissue Segmentation in CT Scan Images

Abstract

Osteoarthritis (OA) is a common cause of painful knee joints among senior citizens. Osteoarthritis restricts the range of movement of the body. Senior citizens are particularly vulnerable, as limited mobility reduces their quality of life, leading to depression and isolation. Proper working of joints is an essential health indicator. In early-stage osteoarthritis can be treated mainly based on medication and physiotherapy. When the conventional treatment stops giving relief to patients, total knee arthroplasty (TKA) is the next option available to the patient. The success of TKA depends on proper sizing and positioning of implant, preoperative and intra-operatively decisions. 3D segmented bone information from the patient�s Magnetic Resonance Imaging (MRI) or Computed Tomography (CT) can help doctors make better pre-operative decisions on implant size and positioning plan. CT scan image is the input for this pre-operative stage of TKA. But the challenge here is that the CT images comes as in slices and it becomes a problem to identify exact location and their physical condition of the bones converging at the knee. The main bones are femur, tibia and fibula. It would become quite helpful for the surgeon if these bones and only bones are visibly available in a 3D view of the given CT scan images. For this segmentation of these bones is desired and if the segmentation be done in an automated way, it would be very less resource and less time consuming. This thesis presents an automated framework for segmentation of the distal femur and femoral head from 3D CT scans. The first stage of this framework consists of 2D U-Net architecture for landmark prediction of the distal femur and femoral head in the original 3D CT scans. The next stage of the framework is to perform local 3D segmentation around the landmark. We also demonstrated and performed tissue and femur bone segmentation using various methods such as fixed threshold segmentation, UNet segmentation, and Ostu�s based segmentation. Further focusing on the femur bone of CT scans of the knee joint, we are using deep learning method the UNet as segmentation model. The test set of CT scans is used to evaluate the methods. The experiments of the results indicate that the method chosen by us, of landmarking and further segmentation is effectively and thoroughly automatically able to segment femur bone and distal femur from the whole leg CT scans.