Quantitative Analysis of Scleroderma
What is Scleroderma?
Scleroderma is a disease which affects many body systems, but is primarily characterized by thickening and tightening of the skin. More women are affected than men, and the dominant age group is between 20 to 50. The lungs are affected in 70-90% of cases, and develop either fibrosis or changes in the blood vessels which lead to increased pressure in the pulmonary arteries. The fibrosis usually starts with with an increase in lung fiber density near the posterior (back) regions of the lungs. CT imagery frequently shows the initial effects as an amorphous increase in lung density with an appearance like "ground glass". Later stages of fibrosis are characterized by the emergence of a network of coarse lines. These coarse lines eventually develop into regions containing large numbers of small cysts. This end-stage effect is sometimes referred to as "honeycombing" and is non-reversable.
The Goal of our Research
The goal of our research is to use High Resolution CT (HRCT) imagery to automatically detect and quantify Scleroderma in its early stages. Scleroderma treatments are most effective when applied during these early stages. Automated detection methods can aid doctors in gauging the effectiveness of new treatments over time. We hope to not only quantify regions of Scleroderma lung tissue, but also qualify them as to their type ("ground glass" density, abnormal lines and "honeycombing").
At present, the status of the disease is assessed visually by a radiologist. However, segmentation of Scleroderma areas by a radiologist can sometimes be a subjective affair. Automated classification routines remove human subjectivity from the measuring process. Removing this subjectivity may allow physicians to more accurately monitor the effectiveness of a particular Scleroderma treatment over time.
The Technique
The automated classification method we have constructed detects Scleroderma abnormalities through the use of a supervised training scheme. In this method, a classifier is "taught" to recognize Scleroderma by feeding it examples of CT image abnormalities. The classifier "learns" from these samples by computing and recognizing relevant features within the training imagery. The current features being used to train and detect Scleroderma include standard deviation, skewness and kurtosis. These three features are computed within small circular neighborhoods around every pixel within the lung image. These local statistical features are then processed through an advanced clustering scheme. The clustering software creates a Scleroderma detection "codebook" by automatically finding a set of clusters in which the statistical features from the training imagery tend to clump about. The detection "codebook" can then be applied to Scleroderma imagery which was not part of the original training set. Our clustering method is based on a Gaussian Maximum Likelihood model in which the clusters are parameterized according to their means and covariances.
Below is an original CT image and two images in which which the Scleroderma areas have have been highlighted in yellow by two different radiologists. The haziness and abnormal lines seen near the edges of the lung are a sign of Scleroderma. Note the differences in the radiologists' interpretation of the Scleroderma areas. Clicking on any of these three images will bring up a higher resolution version.
The image shown below demonstates the results obtain by applying the automated classifiers. The results shown here are well within the bounds of those produced by a trained radiologist. Clicking on this image will bring up a higher resolution version.
Shown below are the four training samples which were used to train the classifier system. Each sample has an inherent set of statistical features that were "learned" by the automated classifier system. These four training samples did not come from the same image which has been classified. They came from a group of several different images that have Scleroderma characteristics which are somewhat similar to the classified image. Clicking on the image below will bring up a higher resolution version.
Automated Segmentation of Scleroderma in High Resolution CT Imagery
Authors: R. Fortson (LANL), Dr. D. Lynch & Dr. J. Newell (NJCIRM)
Adobe PDF Version (473 Kb) sclero.pdf
PRESS HERE to see slide show page on this subject (313 Kbytes).
Consumer information and answers to many questions about various lung diseases can be found by contacting the National Jewish Center for Immunology and Respiratory Medicine in Denver, Colorado.
This work is being performed by the Computer Research and Applications Group at Los Alamos National Laboratory. Please feel free to request additional information or to let us know about related efforts.
Last updated April 27, 1999.