An Associative Memory Applied to Anatomical Analysis of CT Images

We generalize Kohonen's Optimal Linear Associative Memory (OLAM) neural network model and apply it to the analysis of medical images. More specifically, we apply an OLAM to the identification of anatomical structures and abnormalities in CT images.;OLAM is one of the simplest unsupervised learning models, making it a popular alternative to more complicated neural networks models such as back-propagation or the Hebbian model. However, OLAM has several drawbacks. One is that Kohonen's solution to the classical OLAM formulation is unstable, in that small changes in the data accompany large changes in the solution. Another drawback of OLAM is its limited memory capacity, as the dimension of training data vectors is not sufficiently large enough. Still another is its computational complexity, as it requires a large number of floating point operations.;In this thesis, we resolve these problems and apply our new algorithms to the detection of anatomical structures in CT images. We address the stability problem by developing an algorithm for a more stable “nearby” problem, drawing on Bellman's theory of dynamic programming.;We remedy the problems of noisy data and memory capacity of OLAM by choosing training data that is not only less noisy but at the same time increases the memory capacity of OLAM.;OLAM relies on the orthogonalization of data vectors. Unfortunately, the classical Gram-Schmidt orthogonalization process, used by Kohonen, performs poorly when the data vectors are nearly linearly dependent, and completely fails when they are dependent. Our algorithm for the solution to the modified OLAM formulation generalizes the Gram-Schmidt process, producing approximately orthogonal vectors even when the data vectors are nearly dependent or dependent.;We conduct recognition experiments on artificially generated CT images, known as the Shepp-Logan phantoms, of cross sections of the human head. The experiments support our model and perform as expected.;There are several advantages in using the modified OLAM model to complement the diagnostic abilities of clinicians. One is that the model captures the collective experience of many experts, i.e. it can store more diagnostic experience than a single physician, as it does not compromise its abilities due to fatigue or stress. It also quantifies the diagnostic results.