POSTER: METHODS A Neural Net Method To Measure The Volumes Of Brain Regions V.A. Magnotta, D. Heckel, T. Cizadlo, P. Westmoreland Corson, N.C. Andreasen Mental Health ClinicalResearch Center, The University of Iowa College of Medicine, Iowa City, Iowa, U.S.A. Introduction: Neural nets have been proposed as a method that could be applied to solve image analysis problems."! Briefly, neural nets can "Iearn"to identify features of complexstructures, much as the human brain learns to recognize complex patterns. If given adequate information about color, shape, and location (e.g., the gray-matter composition and its anterior location adjacent to the frontal horns and posteriortermination in the amygdala), they can use this information to iteratively check and correctly identify those voxels in the brain that are consistent with this information. We have developeda method to use neural nets as a means to identifyspecific brain structures. Method: The procedure begins with the generation of a training data set, which is produced by manual segmentalion of a given structure by twoexperiencedtechnicians. At present we consider a reliability of .9 (intraclass r) as suitable for training the bet. Segmentation is done on 30 scans which was based on the number of manual tracings that were available. The data from 20 are then used for training, while the remaining 10 are used to check the result. The training process involves scanning all voxels in an individual image data set, one voxel at a time, and determining whether the voxel should be assigned as part of the structure (ROI) identified in the training data set. The network architecture consists of one output node, one hidden layer with approximately 70 nodes, and 47 input nodes. The output node is used as a fuzzy binary value indicating whetherthe voxel is in the ROI or not. The exact number of hidden nodes needed to find a simple ROI such as the caudate was decided on empirically. We tried many different valuesand rated the ability of the trained networkto generalize. The exact numbercan be adjusted so that more are used for complex ROIs and fewer for simple ROIs. The input layer for the network consists of 47 nodes. One is the signal intensity (51) of the voxel being considered by the network and 42 nodesare used to describe the 51 of its surrounding voxels. To provide a computationallyefficient survey of the voxel's neighborhood. we sample a spherical region up to 3 voxels away. using only voxels diagonal and orthogonal to the voxel's location. An additional 3 nodes are used for the 3 dimensions of normalizedTalairach atlas location and one node contains a probabilityestimate. The search space of the networkis determinedby surveying the segmented daIa in approximately 20 subjects. This area is dilated to ensure inclusion of the ROI inall subjects independently of brain size or variability in location. The probability etimate of the input node is given by the frequency with which each location in the search space was found to be inside the structure. The training set consists of a number of input and target output vector pairs based on the hand-traced ROIs. The number of training ROIs may vary from one type of ROllO another, but the number reflects a set of basic principles. The number should he sufficient 10 contain a spectrum of deviances from the "typical" structure of interest. It should be large enough to give a good sample of the different voxel neighborhoods in the search area, but small enough 10 be computationally efficient. Exact location of borders is difficult to trace, so border voxels are weighted less than interior voxels. The vector pairs are then fed into the backpropagation algorithm,and the meansquarederror of rhe net's response, as compared to the target vector from the training set. is tracked over time. The n squared error measurement usually drops off quickly in the early pan of the training and then gradually flattens out at some asymptote. Training is halted when an asymptote is reached. After trainingiscompleted. the quality of the net's identification is then checked. At present, we apply its algorithm to the remaining 10 traced ROls in the initial data set of .'n, We use intraclass R, sensitivity, and specificity to check quality. The adjacent image shows the hippocampus. as identified by this method. To date we have also applied it to the caudate and Heschl's gyrus. Conclusion: Our future work will involve applying this approach to a variety of structures, with the ultimate goal of developing a highly automated and accurate way of measuring all structuresor regions of interest on individual MR scans, thereby making image processing more rapid and reliable. References: I. Raff, U., Scherzinger, A. L.• et al. Physics, 1994; 2)(12): 2. Singer, W. Science, 1995; 270: 758-764. 3. Hinton. G. E. Scientific American,. 1992; 275: 145-151. S697