Automated Knee Bone Segmentation and Visualisation Using Mask RCNN and Marching Cube: Data From The Osteoarthritis Initiative

Abstract

In this work, an automated knee bone segmentation model is proposed. A mask region-based convolutional neural network (RCNN) algorithm is developed to segment the bone and reconstructed into 3D object by using Marching-Cube algorithm. The proposed method is divided into two stages. First, the Mask RCNN is introduced to segment subchondral knee bone from the input MRI sequence. In the second stage, the segmented output from Mask R-CNN is fed as input to the Marching cube algorithm for the 3D reconstruction of knee subchondral bone. The proposed method achieved high dice similarity scores for femur bone 95.35%, tibia bone 95.3%, and patella bone 94.40% using a Mask R-CNN with Resnet-50 as backbone architecture. Improved dice similarity scores for femur bone 97.11%, tibia bone 97.33%, and patella bone 97.05% are obtained by Mask RCNN with Resnet-101 as backbone architecture. It is noted that the Mask RCNN framework has demonstrated efficient and accurate knee subchondral bone detection as well as segmentation for input MRI sequences.