This paper will be presented in the Computational Intelligence, Robotics and Autonomous Systems (CIRAS 03), December 2003. Automated brain data segmentation and pattern recognition using ANN. Carlos A. Parra, Khan Iftekharuddin and Robert Kozma Intelligent Systems and Image Processing Lab Institute of Intelligent Systems University of Memphis. Memphis, TN 38152 Abstract— In this project we implement an artificial neural network (ANN) algorithm to perform the segmentation of brain MRI data. The multispectral characteristics of MR images with different modalities such as T1, T2 and PD are exploited to segment different brain tissues. The ANN algorithm used in this implementation is the Learning Vector Quantization (LVQ) network. The images required for training and test are obtained from a simulated brain database integrated in the McConell Brain Imaging Center (McBIC) of McGill University’s Montreal Neurological Institute. The results of the segmentation algorithms are qualitatively compared to the phantom images to mask each tissue. Our results suggest excellent brain tissue segmentation. We plan to exploit our results in formulating biologically plausible models for automated tumor detection. I. I NTRODUCTION A. Why is MRI a multispectral imaging technique? MRI is an imaging technique based in the measurement of magnetic field vectors generated after an appropriate excitation with strong magnetic fields and radio-frequency pulses in the nuclei of hydrogen atoms present in the water molecules of the patient’s tissues. Given that the content of water differs for each tissue (bone and muscle, for example), it is possible to quantify the differences of radiated magnetic energy, and have elements to identify each tissue. When specific magnetic vec- torial components are measured under controlled conditions, different images can be acquired, and information related to tissue contrast may be obtained, revealing details that can be missed in other measurements. For the present study, I will consider T1, T2 and PD weighted images. (see Fig. 1) (a) T1 (b) T2 (c) PD Fig. 1. MRI multispectral images considered in this study B. What is Segmentation ? The problem of segmentation is central to many image processing applications, and owes its importance to the need of identify or distinguish objects from a background or pinpoint objects embedded in other objects such as in the case of a tumor growing in the Cerebrospinal Fluid (CSF) region. There are four typical approaches segmentation. The Threshold techniques, where the classification of each pixel depends on its own information such as intensity and color information. This technique is efficient when the histograms of objects and background are clearly separated. The Edge-based methods are focused in detecting contour. They fail when the image is blurry or too complex to identify a given border. The Region-based segmentation, in which the concept of extract- ing features (similar texture, intensity levels. homogeneity or sharpness) from a pixel and its neighbors is exploited to derive relevant information for each pixel. A more sophisticated and recent method called Connectivity-preserving relaxation uses spline curves and modifies them (shrink/stretch) apply- ing energy functions. Finally, the Cooperative Hierarchical Computation approach uses pyramid structures to associate the image properties to an array of father nodes, selecting iteratively the points that average or associate to a certain image value. In the specific case of Brain MRI, the problem of segmen- tation is particularly critical for both diagnosis and treatment purposes. In these cases, the accurate location of a lesion is directly related to an early detection of a potential pathology, as well as to minimizing the damage to healthy tissues that can be caused by therapy procedures such as radio-surgery. The brain MRI offers a valuable method to perform pre and post surgical evaluations, which are key to define procedures and to verify their effects. Therefore, it is necessary to develop algorithms to obtain robust image segmentation such that the following may be observed: •Automatic or semiautomatic delineation of areas to be treated prior to radio-surgery •Delineation of tumors before and after surgical or radio- surgical intervention. •Tissue classification: volumes of White Matter, Gray Mat- ter, Cerebrospinal Fluid (CSF), Bone, Muscle (Skin), and Abnormal Tissues. C. Why are ANN’s good for segmentation? MR images are large data sets with an important number of independent variables and complex relationships, that usually show a nonlinear character that makes classical statistical