Automated lesion detection and lesion quantitation in MR images using autoassociative memory

Abstract

Previous efforts concerning lesion extraction in radiologic images indicated that autoassociative memory models can be a valuable tool in automated lesion detection. Preliminary results are expanded to resolve the technical problems of image registration and magnification. Instead of operating on selected portions of the MR images, each entire image matrix is operated upon as image vector comprising all stacked columns of the matrix. Spin density weighted images (TR = 3000 ms and TE = 40 ms) of 42 normal subjects were remapped and standardized with respect to location and magnification. All image vectors were orthonormalized to span a linear manifold. Standardized abnormal image vectors were then tested by the stored autoassociative memory and the abnormalities (novelties) were extracted by application of an autocorrelation matrix to the input vector. The autocorrelation matrix is computed using image vectors from normal subjects. The lesions (multiple sclerosis and tumors) are then identified as the orthogonal component to the linear manifold spanned by the basis vectors of the normal brain scans. Lesion extraction has been achieved with the intention of quantitating and staging diseased parenchyma after automated edge detection.