BCD-NET: A NOVEL METHOD FOR CARTILAGE SEGMENTATION OF KNEE MRI VIA DEEP SEGMENTATION NETWORKS WITH BONE-CARTILAGE-COMPLEX MODELING

摘要

Segmentation of cartilage in knee MRI is an important process for various clinical tasks in diagnosis and treatment planning of osteoarthritis. Recently, the deep segmentation networks (DSNs) have been applied to the cartilage segmentation and have shown promising results. However, the DSNs have limitations for cartilage segmentation in that the networks tend to ignore smaller objects like the cartilage during learning the multi-class segmentation. In this paper, we propose a novel method for DSN-based cartilage segmentation using bone-cartilage-complex (BCC) modeling and bone-BCC-difference (BCD) extraction. To overcome the small cartilage ignorance of DSNs, we propose to construct the BCC which combines bone and cartilage in a single mask. Then we convert the problem of segmenting cartilages into the problem of segmenting the BCCs and bones, and extracting the cartilages by subtracting the bones from the BCCs. In addition, we apply 2.5D segmentation to further improve the segmentation accuracy by averaging the multiple segmentation masks on different planes with majority voting. In experiments, our BCD-Net achieved the average DSCs of 98.1% and 83.8% for femoral and tibial cartilages, respectively, and showed state-of-the-art-level performances in SKI10 public challenge validation dataset.