Anthony H. Butler MBChB, Presenter: Nothing to Disclose

Phil Bones PhD, Abstract Co-Author: Nothing to Disclose

Michael Anthony Hurrell MBChB, Abstract Co-Author: Nothing to Disclose

To investigate dimensionality reduction using eigenimage processing in a classification scheme for chest radiographs.

Eigenimage processing is a statistical technique for reducing high resolution images to a few parameters per image. It was initially developed for face recognition. To date there has been little use for radiological image processing.Eigenimage processing involves starting with a training set of M registered images each of N pixels, where M In statistical terminology the basis images are the principal components of the training set. Mathematically this is the first M eigenvectors of an N x N sized co-variance matrix. For most radiological training sets calculating a co-variance matrix is impractical. To avoid directly calculating such a large matrix we use a method that is closely related to the singular value decomposition (SVD).For our investigation we used a training set of normal frontal chest images (M=77) of a standard size (N~3,000,000 pixels). Chest radiographs were chosen because they use a standardised radiographic technique and are an example of the high resolution images found in radiology.15 new normal chest images and 15 chest images of pneumonia were then parameterised using the previously calculated basis images. The distribution of the parameters was examined by calculating the Euclidean distance from the centre of eigenspace for each image.

The distributions of the distances from centre of eigenspace were different for the normal and pathological images. All the calculations could be performed in a reasonable length of time on a typical personal computer.

Early experience suggests that eigenimage processing can be used for classification of crude pathology. The method shows promise for a wider range of pathology.

Butler, A, Bones, P, Hurrell, M, Eigenimage Processing of High Resolution Chest Radiographs.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4402860.html