[en] In this thesis, we solve inverse radiation transport problems by an Artificial Neural Network(ANN) approach. ANNs have many interesting properties such as nonlinear, parallel, and distributed processing. Some of the promising applications of ANNs are optimization, image and signal processing, system control, etc. In some optimization problems, Hopfield Neural Network(HNN) which has one-layered and fully interconnected neurons with feed-back topology showed that it worked well with acceptable fault tolerance and efficiency. The identification of radioactive source in a medium with a limited number of external detectors is treated as an inverse radiation transport problem in this work. This kind of inverse problem is usually ill-posed and severely under-determined; however, its applications are very useful in many fields including medical diagnosis and nondestructive assay of nuclear materials. Therefore, it is desired to develop efficient and robust solution algorithms. Firstly, we study a representative ANN model which has learning ability and fault tolerance, i.e., feed-forward neural network. It has an error backpropagation learning algorithm processed by reducing error in learning patterns that are usually results of test or calculation. Although it has enough fault tolerance and efficiency, a major obstacle is 'curse of dimensionality'--required number of learning patterns and learning time increase exponentially proportional to the problem size. Therefore, in this thesis, this type of ANN is used as benchmarking the reliability of the solution. Secondly, another approach for solving inverse problems, a modified version of HNN is proposed. When diagonal elements of the interconnection matrix are not zero, HNN may become unstable. However, most problems including this identification problem contain non-zero diagonal elements when programmed on neural networks. According to Soulie et al., discrete random iterations could produce the stable minimum state of an associative memory. We modify the conventional HNN into a new HNN which has random delayed updating intervals in order to alleviate the above unstable phenomenon. So we shall call it 'Delayed Hopfield-Like Neural Network(DHNN).' We tested DHNN under various noisy environments to verify its efficiency and noise robustness, and compared with previous works in this category. It proved that our approach is more robust and efficient than previous ones in a number of tests. Although the probability of successful identification decreases as the noise level increases, the successful identification rate is still acceptable. Further, we investigated and applied DHNN to some potential applications, such as medical imaging (tumor or cancer detection) and nuclear waste assay, and showed that these problems could be solved by our approach. As a concluding remark, we may say that DHNN can solve ill-posed problems with reasonable efficiency and robustness, including the problem of identification of the radioactive source in a medium