Abstract: An improved fuzzy CMAC neural network，called IFCMAC，is presented.The difference between the classical FCMAC and IFCMAC is that linear weighted sum of input vector is introduced between fuzzy after connected layer and output layer to compensate for approximation error.Then its approximation accuracy is enhanced and the common fault of the series of CMAC neural network can be solved.In the simulation for the segmentation of the head magnetic resonance images，the gray value of the current pixel and texture features of neighborhood of the current pixel are used as the inputs for IFCMAC.The simulation also demonstrates that its effect of segmentation is better than the classical FCMAC.