الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 125 |  |  |
| | | from | 125 | | |
Abstract
The knee is an important part in the human body. It helps a person to move, it is also one of the complex joints. By virtue of the anatomical nature of the joint, it has a weak composition compared to its load on the human body. It is vulnerable to injury with less force than the impact, such as making a sudden violent movement or carrying excessive weight on the knee joint or its constituent bones. Numerous issues can affect the knee, most notably the meniscus of knee, the cruciate ligament, knee roughness and knee cancer. Early prediction of knee diseases contributes greatly to preserving them, and thanks to advancements in machine learning and visual computing, it has become possible to use them to predict the occurrence of any problem in the knee and then maintain it. In this dissertation , the to make the most recently artificial intelligence techniques, represented by deep learning algorithms combined with computer vision had been used. The field of computer vision is a subset of AI that allows for the automatic analysis of visual data such as images and movies.. Computer vision, likes artificial intelligence, allows computers to observe, watch, and comprehend in the same way that people do. Deep learning techniques have become extremely popular and widely employed in computer vision in recent years due to their ability to extract attributes. Convolution Neural Network is an advanced deep learning approach that excels at image categorization. For the purpose of segmenting biological images, the University of Freiburg developed an improved version of CNN in 2015 called U-Net.
Deep Learning and Computer Vision Techniques are utilized to diagnose common knee problems automatically. In this dissertation, AI methods like deep learning and computer vision had been applied to diagnose common knee problems automatically. The knee problems discussed in this dissertation are the meniscus torn in the knee. The cruciate ligament tears , knee osteoarthritis and knee bone tumor.
In this dissertation four different work would be presented for common knee problems. The first work, Knee Images Classification using Transfer Learning was introduced automatic model for meniscus torn and anterior cruciate ligament tears classification. A pre-trained model based on deep transfer learning was used. The data set used in this work is hug, 45200 images that collected by Stanford ML University. The pre-trained NAS-Net Mobile model used to extract the image features. Four different classifiers were used. The random forest achieves the maximum accuracy for all planes with average accuracy 99.4%, 98.1% and 97.6% for knee abnormal disease, meniscus and anterior cruciate ligament respectively, NN gives the worst accuracy for all planes with average accuracy 73.86%, 58.56% and 60.16% for knee abnormal disease, meniscus and anterior cruciate ligament, respectively.
The second work introduced model “Automatic Knee Anterior Cruciate Ligament Torn Diagnosis Using CNN-XGBoost ” for anterior cruciate ligament tears classification. 130 MRI knee images were used. To extract the ACL region from the knee picture, the region growing based segmentation technique was utilized, and then the CNN mixed with the XG-Boost model was applied for knee anterior cruciate ligament classification. The overall system gives accuracy 91% and area under the curve equal 0.92.
The third work “Bone Osteosarcoma tumor classification” introduced automatic model for Knee Bone Osteosarcoma tumor classification. 1091 histology images stained with hematoxylin and eosin. (H&E) that collected by The Dallas-based institution University of Texas Southwestern was utilized. The images are classified into three different classes, tumor tiles, osteosarcoma, and non-tumor images. The color normalization are first performed to improves the quality of digital tissue slides and increases perceived image quality while generating few artefacts. After that the color de-convolution is performed to separate original images into haematoxylin, eosin images. The proposed approach uses the gray level co-occurrence matrix to extract texture features for pathology images. The proposed uses three different classifier for feature extraction, XG-Boost, support vector machine, and k-nearest neighbors. Finally, ensemble voting is used by combining the predictions from these classifiers. The proposed system gives 91.8% accuracy.
The fourth work introduced model for knee osteoarthritis automatic detection. 500 X-ray images were collected by the National Institutes of Health (NIH) are used. These data classified Knee OA as stated by Kellgren and Lawrence classification, from grade zero to grade four. The image is annotated using VIA tool to create ground truth image. The proposed system uses the U-Net architecture to automatically recognize the OA grade and classify it as stated by Kellgren and Lawrence, which classify the knee osteoarthritis to four different grade from zero to four. Three different metrics are used to evaluate the quality of segmentation, intersection over union, pixel identity, and localization error. The overall accuracy of the method was 96.3%. The suggested model computes the JSW between the tibia and femur bones automatically. U-Net is a popular network design used in the suggested model. The suggested model is applicable to five classes of OA and achieves average (IOU,PI,LE) values of (0.8448, 0.9549, 5.322), (0.8, 0.948, 7.474), (0.826, 0.952, 4.319), (0.761, 0.948, 7.4696), and (0.748, 0.924, 20.251) for grades 0, 1, 2, 3, and 4.