Automated Knee Articular Cartilage Segmentation Using Convolutional Neural Network (CNN): Preliminary Results

Automated identification and segmentation of medical imaging data is very desirable. The reason is typically the large size of data and therefore enormous time, which radiologist has to invest in manual segmentation. Therefore, we developed a simplified convolutional neural network (CNN), which could efficiently and automatically segment the knee articular cartilage into multiple classes on a currently available graphical processing unit (GPU). Currently available dataset of ten manually segmented patients MRIs divided into three subsets were used for training, validating and testing our neural network. Therefore our results are only preliminary but fulfill our expectations. Results could be more precise by involvement data augmentation step in training CNN process, which will be realized by our team in the near future. We continually work on the enlargement of our dataset, as well as on the involvement of bigger available datasets from other scientific groups.