Inspection and shape measurement of three-dimensional objects are widely needed in the fields of quality monitoring and reverse engineering. X-ray computed tomography could be a good solution since the method can acquire three dimensional volume information of a product from a series of acquired cross-sectional images. To reconstruct a cross-section in computed tomography, a number of data are required, projected from all but uniformly spaced view angles. In many applications of industrial field, however, it may not be possible to acquire such projection data obtained at all angles due to the size of objects or obstructed situation by other structures at some angles. In such a limited condition, analytical solution to reconstruct a cross-section is not available in general, and an iterative numerical method such as algebraic reconstruct technique (ART) and its modified algorithms, such as uniform and simultaneous ART methods, are used. In those iterative methods, the intensities of the image pixels in the reconstructed image are estimated and updated independently, thus the reconstructed image looks like a mosaic, of which the resolution is restricted to the number of image elements, pixels. In this paper, a new image reconstruction method is proposed based on a radial basis f function (RBF) neural network. In this method, a cross-section image is represented by RBF network, and is reconstructed through the learning process of the network. To achieve this, a learning method of the network is proposed here based on the projection of the image instead of the reference image itself. The algorithm is tested by a series of simulation studies on some of modeled images, and the performance of the proposed method is evaluated in terms of convergence and accuracy.