Automatic algorithm for knee cartilage segmentation on quantitative MRI images

Algorithm

The algorithm is based on several segmentation methods. Initially a bone segmentation is done by Watershed segmentation and morphological operations. The bone segmentation allows for slice specific mask generation to only allow cartilage to be segmented at anatomically valid locations. The segmentation within the masks is then done by a k-means classification algorithm, customized to quantitative MRI knee images.

Data

Quantitative MRI allows for meaningful investigation on physical and chemical variables in physical units and also comparison between tissue regions on different patients as well as between different time occasions. This since Quantitative MR imaging provides an absolute MRI scale in images. Data sets where provided by SyntheticMR.

Processing steps

Watershed segmentation and Morphological operations

Dilated bone contour mask

Positioned cartilage mask

K-means classification

Results

The algorithm was evaluated on the test patient given by SyntheticMR on the inner slices. The final algorithm reached a recall of 0.8394, a precision of 0.2835 and a F1score of 0.4238.

Comments

The algorithm is far from ready for clinical use, however the results from the first attempt at segmenting cartilage shows promise. In the technical documentation suggestions can be found of future actions.

Documentation

The project resultet in the following documentation:

Technical documentation

Also, here's a video of a funny cat:

Group Members

FIVE Biomedical Engineering students from Linköping University faced a cartilage segmentation challenge. A challenge provided by ONE company, i.e. SyntheticMR.

Emma Riedl

Justus Johansson-Lindkvist

Maria Kastberg

Rasmus Dencker

Simon Karlsson