Abstract Title
Automated Segmentation of the Knee Using a Convolutional Neural Network
Additional Funding Sources
This work was supported by the National Institute on Aging Grant R15 AG059655.
Abstract
For this project an algorithm was developed that is able to successfully segment the knee which includes the femur, tibia, and patella along with their respective cartilage structures. This algorithm is currently being updated to include the meniscus which is found within the knee. The algorithm was trained on the Osteoarthritis Initiative (OAI) dataset which consisted of around 20 patients with a variety of different quality MRIs. Each patient dataset that we used to train our algorithm on had between 30 to 160 sagittal images which brought the total number of training images to 3583. Thus far, the algorithm includes six different trained models, one for each structure defined above—the meniscus will be added as the seventh model. Now that we have trained models that are able to segment the knee, we can import a set of new sagittal images to the algorithm and it will start to predict and “stack” these slices in order to create a 3D representation of the knee. These 3D representations will ultimately be used for subject-specific finite element simulations to evaluate joint mechanics.
Automated Segmentation of the Knee Using a Convolutional Neural Network
For this project an algorithm was developed that is able to successfully segment the knee which includes the femur, tibia, and patella along with their respective cartilage structures. This algorithm is currently being updated to include the meniscus which is found within the knee. The algorithm was trained on the Osteoarthritis Initiative (OAI) dataset which consisted of around 20 patients with a variety of different quality MRIs. Each patient dataset that we used to train our algorithm on had between 30 to 160 sagittal images which brought the total number of training images to 3583. Thus far, the algorithm includes six different trained models, one for each structure defined above—the meniscus will be added as the seventh model. Now that we have trained models that are able to segment the knee, we can import a set of new sagittal images to the algorithm and it will start to predict and “stack” these slices in order to create a 3D representation of the knee. These 3D representations will ultimately be used for subject-specific finite element simulations to evaluate joint mechanics.
Comments
W95