||Analysis of Mouse Joints - Examination of Osteoarthritis by Automatic Visual Inspection
||Ersbøll, Bjarne Kjær (Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
||Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
||This project examines the use of automatic image processing to test for osteoarthritis (OA) in laboratory mice. By using mice that are predisposed for developing osteoarthritis, there should be a high correlation between their age and their osteoarthritis stage. The purpose of this project is to generate an automatic OA measure of the osteoarthritis stage of each mouse. Visiopharm has earlier worked on the project and would like to find out if the results can be improved by other approaches than the ones they have chosen. They mainly calculate features from histograms of RGB and other color representations. Their obtained OA measure has a highly significant correlation with age at 0.54.
The data for this project consists of images of half of the right knee (the medial side of the tibia seen from above) of the mice at different ages. The mice are sacrificed and their tibia removed and stained with a blue dye. The dye binds to proteoglycan (an active part of the cartilage) that is destroyed by osteoarthritis. It is the belief that the healthy parts of the tibia appear blue in the images, the first sign of lesion appears purple and the worse lesion areas are bright / white. By automatic image processing, it is attempted to identify the different types of areas and their relative amounts are used to calculate an OA measure. This OA measure is tested against age and other measures known about the mice.
There is generated a probability map of the lesions and it is thereby shown that the lesions mainly emerge on or near the upper border of the tibia. Average images of the tibias at different ages are generated to see how the different lesion types behave according to age in different positions of the tibia. The purple areas are shown to be more concentrated near the bright lesions and hence are likely to represent an early lesion stage.
Samples from the images reveal that the different types of areas are not grouped but are mixed to an oblong point cloud in feature space. There is therefore not a sharp border between the different types of areas and that makes their separation and identification more difficult. Eight types of areas are defined and labelled in the images and samples from these are used throughout the report. It is the basic types of areas (blue, purple and white) and subdefinitions of these that give the eight classes. Examination of the classes shows that by looking at the images separately, the classes are pretty much separable but merging the class' centers across the images results in overlapping classes. Thus there is a difference between the images which is found to be somewhat systematic for all the classes in the images and therefore it might be possible to reduce it. Trials with different color transformations cannot remove the difference but only change and improve it for some classes and worsen it for others, depending on which color transformation is used. Among others, trichromatic colors and the IHS transformation are used.
Due to non-separable groups in feature space, clustering is not believed to be the perfect solution and likewise with classification due to the overlapping classes. Both methods are tried. Classification is tried first, to learn more about the classes and the difficulties in the images, and hence improve the knowledge for a clustering solution.
The classification of the images uses different combinations of input and output, to see the effect of the different classes. The classification is tried with each of the colorbands separately, and up to three of them simultaneously. From the implemented Bayes classifier, using Mahalanobis distance, the highest correlation of the OA measure with age is 0.58. Due to possible collinearity, a more reliable result is 0.56, obtained using RGB as input. The results are based on all eight defined classes. Merging the classes to the three basis classes, results in an OA measure that correlates with age at 0.57, also with RGB as input. A set of manually defined decision rules are tried in order to classify the three basis classes. Here, the OA measure's correlation with age is 0.44 and hence not an improvement.
A couple of clustering approaches are suggested and a clustering survey is carried out. Trials do not separate the point cloud in feature space into directly usable clusters. Blue and white overlap and likewise with blue and purple. Strong tendencies according to age are not found and further actions to use the obtained clusters would be similar to the manually defined decision rules, tried in classification. Clustering is therefore not believed to be a usable approach for these images.
The classifier is tried again and improved by using a priori knowledge about the position of the bright lesions. Further improvements are removal of collinearity and noise reduction, among others. The noise is reduced using a median ¯lter and this improves the classification. The obtained correlation is 0.60, using RGB as input. The addition of position information results in an improved correlation at 0.66, which is the maximal obtained result during this project. The interpretation of the last optimization is somewhat unclear hence the 0.60 is the most correct result.
In this project the purple areas are found to be an early lesion stage and the bright / white areas are found to mainly emerge near the upper border of the tibia. It is shown that automatic image processing can be used to establish a reliable OA measure. The classifier approach obtains results at least as good as Visiopharm's earlier solution by measures on histograms. It is believed that the classifier solution can be optimized even further.
||Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU : DK-2800 Kgs. Lyngby, Denmark
||Image analysis; Classification; Clustering; Clustering survey; Color transformation; Osteoarthritis; blue staining; laboratory mice - Billedanalyse; klassifikation; farvetransformation; slidgigt; indfarvning af skinneben; laboratoriemus
Creation date: 2006-06-22
Update date: 2012-12-20