Fully automated segmentation of wrist bones on T2-weighted fat-suppressed MR images in early rheumatoid arthritis

Lun Matthew Wong, Lin Shi, Fan Xiao, James Francis Griffith


Background: Magnetic resonance imaging (MRI) allows accurate determination of soft tissue and bone inflammation in rheumatoid arthritis. Inflammation can be measured semi-quantitatively using the well-established RA-MRI scoring system (RAMRIS), but its application is time consuming in routine clinical practice. To fully realize an automated quantitation of inflammation scoring for clinical use, automatic segmentation of the wrist bones on MR imaging is needed. Most previous studies extracted the wrist bones on T1-weighted (T1W) MR images, and then used registration to segment T2W fat-suppressed images for bone marrow oedema quantification, introducing spatial errors into the process. Relatively little work has tried segmentation directly from T2W fat-suppressed images and no prior study have used convolution neural network (CNN) to segment the wrist bones. The purpose of this study is to develop a CNN-based algorithm for automated segmentation of the wrist bones in early rheumatoid arthritis (ERA) patients on T2W fat-saturated MR images.
Methods: As preliminary tests indicated that out-of-the-box segmentation CNN U-net performance was compromised by close apposition of wrist tendons and bone, we separated the volumes prior to segmentation by using classification CNN Inception V3 to group images with similar features. The classified images were then segmented by individually trained U-net. We trained the networks on 40 cases and tested them on 11 cases derived from an MR imaging dataset of 51 patients with varying severity of ERA.
Results: We obtained a wrist bone segmentation with an average dice similarity coefficient (DICE) of 0.888±0.014, when compared to a manually drawn label. These results are comparable to existing atlas-based methods.
Conclusions: We have developed a fully automatic method to segment the wrist bones in ERA patients of varying severity directly from T2W fat-suppressed MR images. This compares well with manually drawn labels.