Segmentation of knee MRI data with convolutional neural networks for semi-automated three-dimensional surface-based analysis of cartilage morphology and composition

Abstract

To assess automatic segmentations for surface-based analysis of cartilage morphology and composition on knee magnetic resonance (MR) images. 2D and 3D U-Nets were trained on double echo steady state (DESS) images from the publicly available Osteoarthritis Initiative (OAI) dataset with femoral and tibial bone and cartilage segmentations provided by the Zuse Institute Berlin (ZIB). The U-Nets were used to perform automatic segmentation of femoral and tibial bone-cartilage structures (bone and cartilage segmentations combined into one structure) from the DESS images. T 2 -weighted images from the OAI dataset were registered to the DESS images and used for T 2 map calculation. Using the 3D cartilage surface mapping (3D-CaSM) method, surface-based analysis of cartilage morphology (thickness) and composition ( T 2 ) was performed using both manual and network-generated segmentations from OAI ZIB testing images. Bland-Altman analyses were performed to evaluate the accuracy of the extracted cartilage thickness and T 2 measurements from both U-Nets compared to manual segmentations. Bland-Altman analysis showed a mean bias [95% limits of agreement] for femoral and tibial cartilage thickness measurements ranging between -0.12 to 0.33 [-0.28, 0.96] mm with 2D U-Net and 0.07 to 0.14 [-0.14, 0.39] mm with 3D U-Net. For T 2 , the mean bias [95% limits of agreement] ranged between -0.16 to 1.32 [-4.71, 4.83] ms with 2D U-Net and -0.05 to 0.46 [-2.47, 3.39] ms with 3D U-Net. While both 2D and 3D U-Nets exemplified the time-efficiency benefit of using deep learning methods for generating the required segmentations, segmentations from 3D U-Nets demonstrated higher accuracy in the extracted thickness and T 2 features using 3D-CaSM compared to the segmentations from 2D U-Nets.