Knee Bone Segmentation on Three-Dimensional MRI

Abstract

Three-dimensional (3D) images are widely used in the medical field (e.g., CT, MRI). In osteoarthritis research, 3D magnetic resonance imaging (MRI) provides a noninvasive way to study soft-tissue structures including hyaline cartilage, meniscus, muscle, bone marrow lesion, etc. The measurement of those structures can be greatly improved by accurately locating the bone structure. U-net is a convolutional neural network developed for biological image segmentation using limited training data. The original U-net takes a single 2D image as input and generates a binary 2D image as output. In this paper, we modified the U-net model to identify the bone structure on 3D knee MRI, which is a sequence of multiple 2D slices. Instead of taking a single image as input, the modified U-net takes multiple adjacent slices as input. The output is still a single binary image which is the segmentation result of the center slice in the input sequence. By using 99 knee MRI cases, where each knee case includes 160 2D slices, the proposed model was trained, validated, and tested. The dice coefficient, similarity, and area error metrics rate were tallied to assess the performance and the quality of the testing sets. Without any post-processing of the images, the model achieved promising segmentation performance with the Dice coefficient (DICE) 97.22% on the testing dataset. To achieve the best performance, diverse models were trained using different strategies including different numbers of input channels and different input image sizes. The experiment results indicate that the incorporation of neighboring slices generated better segmentation performance than using the single slice. We also found that a larger image size (uncompressed) corresponds to better performance. In summary, our best segmentation performance was achieved using five adjacent neighbor slices (two left neighbors + two right neighbors + the center slice) with the original image size of 352 × 352 pixels.