Fully automated subchondral bone segmentation from knee MR images: Data from the Osteoarthritis Initiative. (1st September 2017)

• Record Type: Journal Article
• Authors: Gandhamal, Akash; Talbar, Sanjay; Gajre, Suhas; Razak, Ruslan; Hani, Ahmad Fadzil M.; Kumar, Dileep
• Is Part Of: Computers in biology and medicine. Volume 88(2017)
• Page Start: 110
• Page End: 125
• Publication Date: 2017-09-01
• Subjects: Subchondral bone segmentation -- Magnetic resonance imaging -- Gray-level S-curve transformation -- 3D multi-edge overlapping -- Boundary correction -- Boundary displacement; Medicine -- Data processing -- Periodicals; Biology -- Data processing -- Periodicals; 610.285
• DOI: 10.1016/j.compbiomed.2017.07.008
• Languages: English
• ISSNs: 0010-4825

Abstract

Abstract: Knee osteoarthritis (OA) progression can be monitored by measuring changes in the subchondral bone structure such as area and shape from MR images as an imaging biomarker. However, measurements of these minute changes are highly dependent on the accurate segmentation of bone tissue from MR images and it is challenging task due to the complex tissue structure and inadequate image contrast/brightness. In this paper, a fully automated method for segmenting subchondral bone from knee MR images is proposed. Here, the contrast of knee MR images is enhanced using a gray-level S-curve transformation followed by automatic seed point detection using a three-dimensional multi-edge overlapping technique. Successively, bone regions are initially extracted using distance-regularized level-set evolution followed by identification and correction of leakages along the bone boundary regions using a boundary displacement technique. The performance of the developed technique is evaluated against ground truths by measuring sensitivity, specificity, dice similarity coefficient (DSC), average surface distance (AvgD) and root mean square surface distance (RMSD). An average sensitivity (91.14%), specificity (99.12%) and DSC (90.28%) with 95% confidence interval (CI) in the range 89.74–92.54%, 98.93–99.31% and 88.68–91.88% respectively is achieved for the femur bone segmentation in 8 datasets. For tibia bone, average sensitivity (90.69%), specificity (99.65%) and DSC (91.35%) with 95% CI in the range 88.59–92.79%, 99.50–99.80% and 88.68–91.88% respectively is achieved. AvgD and RMSD values for femur are 1.43 ± 0.23 (mm) and 2.10 ± 0.35 (mm) respectively while for tibia, the values are 0.95 ± 0.28 (mm) and 1.30 ± 0.42 (mm) respectively that demonstrates acceptable error between proposed method and ground truths. In conclusion, results obtained in this work demonstrate substantially significant performance with consistency and robustness that led the proposed method to be applicable for large scale and longitudinal knee OA studies in clinical settings.