Fully automated segmentation of cartilage from the MR images of knee using a multi‐atlas and local structural analysis method

Abstract

Purpose: To develop a fully automated method to segment cartilage from the magnetic resonance (MR) images of knee and to evaluate the performance of the method on a public, open dataset. Methods: The segmentation scheme consisted of three procedures: multiple-atlas building, applying a locally weighted vote (LWV), and region adjustment. In the atlas building procedure, all training cases were registered to a target image by a nonrigid registration scheme and the best matched atlases selected. A LWV algorithm was applied to merge the information from these atlases and generate the initial segmentation result. Subsequently, for the region adjustment procedure, the statistical information of bone, cartilage, and surrounding regions was computed from the initial segmentation result. The statistical information directed the automated determination of the seed points inside and outside bone regions for the graph-cut based method. Finally, the region adjustment was conducted by the revision of outliers and the inclusion of abnormal bone regions. Results: A total of 150 knee MR images from a public, open dataset (available at www.ski10.org) were used for the development and evaluation of this approach. The 150 cases were divided into the training set (100 cases) and the test set (50 cases). The cartilages were segmented successfully in all test cases in an average of 40 min computation time. The average dice similarity coefficient was 71.7% ± 8.0% for femoral and 72.4% ± 6.9% for tibial cartilage. Conclusions:The authors have developed a fully automated segmentation program for knee cartilage from MR images. The performance of the program based on 50 test cases was highly promising.

Automatic segmentation of high- and low-field knee MRIs using knee image quantification with data from the osteoarthritis initiative

“…Some methods are directly hierarchical starting with a bone segmentation that then aids the (more difficult) cartilage segmentation with features for distance-from-bone or position-relative-to-bone. 18,36,[38][39][40][41][42] Other methods achieve a similar coarse-to-fine effect by applying atlas-based registration before either nonrigid registration 21,40,43 and/or classifier-based segmentation. 10,22,43 In general, it appears that integration of global information and local features is essential for solving the challenging problem of segmentation of cartilage on the background of multiple other tissue types.…”

“…It would appear that dedicated methods are needed for robust and accurate joint registration-such as articulated-rigid bone registration followed by locally elastic registration in cartilage areas. The third-ranking SKI10 method actually applies a multiatlas nonrigid registration approach to produce an initial segmentation that is used as an input to a graph-cut method, resulting in cartilage overlap errors of 28.3% and 27.6%, corresponding to Dice 0.677 and 0.719 for tibial and femoral cartilages (from UPMC, previous version presented 21 ). The fifth-ranked method 22 also applies multiatlas registration using patch-based label fusion (previously demonstrated to be very applicable for brain MRI segmentation) followed by a specialized three-class classification.…”

Multiatlas Segmentation Using Robust Feature-Based Registration

“…the STAPLE algorithm [34] and different types of machine-learning approaches [27]. The fused segmentation proposal can be further refined into a final segmentation by using graph cut [3,22] or random forest-based methods [9]. For a comprehensive survey of multiatlas segmentation methods and their applications, see [13].…”

Unsupervised segmentation of MRI knees using image partition forests

“…The most successful methods so far have been based on active appearance models, 5 statistical shape models and graph-based optimization, 6 and multi-atlas and local structural analysis. 7 Regardless of the specifics, a common characteristic of all methods is their processing MRI images as arrays of pixels.…”

“…Recent literature [4][5][6][7] reports several attempts in automatic segmentation of the knee structures (mainly femur, tibia and cartilage) with various degrees of success. The most successful methods so far have been based on active appearance models, 5 statistical shape models and graph-based optimization, 6 and multi-atlas and local structural analysis.…”