Comparison and classification of images obtained from a single or more patients, at different times but with the same procedure, is important to evaluate the origin or the degree of several pathologies. As well, image classification fusing data acquired from different sources is often needed to locate regions or volumes, to analyze complex scenes or to simulate a diagnosis prediction. In this paper we present an enhanced neural model able to locate and classify tissue densitometric alterations in CT/MR image sequences.The model is based on a three-step procedure: a) registration step; b) features extraction step; c) classification step. In the first step a neural network based approach is used to match the slices with a reference image in order to become comparable. In the second step a co-occurrence matrices based approach is used to extract the main features from the images and, in the end, the extracted features are classified by mean of a Self Organizing Map.The high computational complexity of this model has been specifically optimized in order to reduce the execution time. The registration step has been optimized at two different levels: by introducing an efficient method for corresponding pixels detection and by performing a parallel implementation of the algorithms. The features extraction step has been optimized by introducing look up tables in order to drastically reduce the number of co-occurrence matrices to compute. The classification step does not need a specific optimization since the learning phase is offline performed.The algorithms implemented have been tested on real complex cases: a database of about 200 CT/MRI brain and abdominal slices was used relative to 40 different patients, with the resolution of 512x512 pixels. Both normal and pathologic conditions (such as cancer or hematoma), have been processed in order to characterize the lesions and to derive the image classification.Other optimization approaches are being analyzed in order to further reduce the computational complexity and to avoid the human assisted and image dependant procedures. Future extension of this model should consider a full 3D modeling of the images and the comparison and characterization of the tridimensional models themselves.