Abstract:
"Knee injury is a common and frequently incapacitating problem that affects people of all ages,
genders, and lifestyles. The knee joint is particularly sensitive to damage because of its
complexity and the pressures that sports, vocations, and daily activities use on it. The most
effective way to identify a knee injury is through magnetic resonance imaging (MRI).
Understanding a knee MRI is a time-consuming and highly sensitive task. Depending on the
radiologist's experience and skill, error rates can vary. To overcome those issues, the author
proposed an automated knee MRI classification system using computer vision.
Computer vision models like ResNet, AlexNet, VGG, and DenseNet are commonly used for
knee MRI classification tasks. However advanced computer vision models InceptionNet are
rarely used for knee MRI classification tasks. Some previous studies were used only for one type
of knee injury. In this research, the author proposed a new model using InceptionNet for three
types of knee injuries (ACL, meniscus, and knee abnormality).
The author trained nine models, corresponding to three injuries and three planes using
InceptionNet and modified the layers of the architecture with the pre-trained weights from
ImageNet. The author finally achieved an accuracy ranging from 75% to 95% accuracy for each
model."
