Neural network object recognition for inspection of patient setup in radiation therapy using portal images

This paper presents a method to improve the performance of inspection of patient setup in radiation therapy using portal images. The proposed algorithm is used to match real-time portal and simulator images which contain the information of patient position relative to the beam position. Two approaches are proposed for this purpose. One approach is to use a multi-scale (multiresolution) analysis of the acquired image where the image is examined at several levels of detail simultaneously. Another approach is to use a more robust optimization method to find the global minimum of a cost function which is generalized from matching both anatomical structures and radiation treatment fields. The optimization method used here is a concurrent (coarse-and-fine) multiresolution model-based object recognition technique using a multi-layer Hopfield neural network. The performance of the algorithm is demonstrated in a clinical radiotherapy application.