Neural network reconstruction of single-photon emission computed tomography images

Abstract

Ah artificial neural network (ANN) trained on highquality medical tomograms or phantom images may be able to learn the planar data-to-tomographic image relationship with very high precision. As a result, a properly trained ANN can produce comparably accurate image reconstruction without the high computational cost inherent in some traditional reconstruction techniques. We have previously shown that a standard backpropagation neural network can be trained to reconstruct sections of single photon emission computed tomography (SPECT) images based on the planar image projections as inputs. In this study, we present a method of deriving activation functions for a backpropagation ANN that make it readily trainable for full SPECT image reconstruction. The activation functions used for this work are based on the estimated probability density functions (PDFs) of the ANN training set data. The statistically tailored ANN and the standard sigmoidal backpropagation ANN methods are compared both in terms of their trainability and generalization ability. The results presented show that a statistically tailored ANN can reconstruct novel tomographic images of a quality comparable with that of the images used to train the network. Ultimately, an adequately trained ANN should be able to properly compensate for physical photon transport effects, background noise, and artifacts while reconstructing the tomographŸ image. structions will be of a quality comparable to the quality of the ANN training set images.Once the ANN is trained for image reconstruction, novel planar data need only be fed forward through the network to quickly generate the reconstructed image at the outputs of the ANN. However, backpropagation training of a neural network is a rather inefficient descending-hiU search algorithm. 23,24 The inefficiency in training an ANN stems from the thousands or often several thousands of iterations that the training set data must be presented to the network before it sutficiently learns the input-output relationship. Consequently, while a trained ANN can perform image reconstructions quickly and simply, the one-time training process itself can be very time consuming and may be impractical when very large network architectures are used, or problems requiring large training sets are attempted.Fortunately, we have taken advantage of the parallel nature of the neural network by implementing the ANN on a parallel computer. The MasPar MP-2 is a single-instruction multipledata massively parallel system that is composed of a 64 x 64 interconnected mesh of processing elements (PEs). The functions of nodes in each layer of an ANN distributed on the MP-2 can be executed simultaneously by dedicating a PE to execute the processes of one node in each layer of the network. This greatly improves the time in which an iteration of the ANN training set, or training epoch, can be performed.In previous work, we and others have shown that image reconstruction with a backpropagation ANN is feasible, 25,26 and that better generalization and faster trai...

A Perspective on Deep Imaging

“…There have been a number of initial attempts at using machine learning and deep learning for medical image reconstruction. With the use of neural network, two classical papers from more than 20 years ago targeted SPECT image reconstruction [28,29]. More recently, dictionary learning, which is a contemporary machine learning approach, was adapted for MRI and CT image reconstruction [27,30,31].…”

Tomographic Reconstruction: General Approach to Fast Back-Projection Algorithms

“…The development of neural network-based approaches to tomographic reconstruction began in 1995 with the landmark work by Kerr et al [73]. They demonstrated that a neural network could produce tomographic images of comparable accuracy to those used in its training.…”

A New Approach to Image Reconstruction from Projections Using a Recurrent Neural Network

“…In the framework of soft computing, the most popular approach to image reconstruction from projections is based on neural networks-a very popular and important tool of artificial intelligence systems for solving image processing problems, e.g., as described in (Cierniak and Rutkowski, 2000). The idea of a neural network applied to image reconstruction from projections is presented in (Kerr and Barlett, 1995a;Kerr and Barlett, 1995b;Kerr and Barlett, 1995c;Knoll et al, 1999;Munllay et al, 1994). Unfortunately, the supervised learning of algorithms described in these papers cannot lead to good performance.…”