Automated Segmentation of Knee Articular Cartilages in 3D MRI Data of Human Patients: Using Deep Learning Approach

An effective convolutional neural network (CNN) with 3D U-Net architecture for automated segmentation of anatomical-morphological structures in high-resolution 3-dimensional double-echo steady state (3D-DESS) magnetic resonance imaging (MRI) data of human knee joint was proposed. The CNN was trained and validated on freely available manually segmented OAI-ZIB dataset of 507 subjects using high-performance computing (HPC) resources. Obtained CNN model for automated segmentation of bones and cartilages in knee-joint MRI data of subjects in testing subset of the dataset is in comparison with a currently available alternative Model-based segmentation algorithm from 2.6 to 7.1% more accurate (Dice similarity coefficient as a metric) with significantly shorter inference time per subject on currently available GPU accelerated personal computer.