Developing a new approach for material discrimination using modular radial basis neural networks based on dual-energy X-ray radiography

Abstract

Challenges, especially, identification of correlation of attenuation coefficients at high/low energy X-ray, stochastic noise on images, source energy fluctuation, and angular drop-off of X-ray energy/dose rate affect the quality of the material discrimination using dual-energy X-ray radiography. In this study, a new approach using radial basis network (RBN) is applied for the identification of materials. A modular structure, in which each module is responsible for the identification of a single material-thickness is developed. Each module with its specific RBN not only is trained to identify its own material-thickness but also is trained to reject the other ones. The activation/radial function of each network is with standard deviation proportional to the distribution of a specific material-thickness transparency. This technique tries to handle the effects of the mentioned challenges which have quasi-radial distribution on the constructed image. Moreover, the cross-correlation between attenuation coefficients at high/low energy X-ray is detected by the neural network. The results show an acceptable performance of the proposed approach. The noticeable advantages are: 1- Instead of utilizing qualitative discrimination techniques such as pseudo-colouring, the materials are discriminated quantitatively which is more important when the number of material- thicknesses is increased. 2- The modular structure makes it possible to add new material- thickness without unfavourably affecting the existing networks (i.e., there is no need to retrain them). 3- The exact percentage of correct identification, incorrect identification, and misidentification are given 4- Bipolar representation makes possible to identify material-thickness without need/with less need to human operators 5- Material discrimination may be done by a more straightforward computational tool because classical machine learning is applied instead of deep learning.