The use of artificial neural networks in classifying lung scintigrams
An introduction to nuclear medical imaging and artificial neural networks (ANNs) is first given. Lung scintigrams are classified using ANNs in this study. Initial experiments using raw data are first reported. These networks did not produce suitable outputs, and a data compression method was next employed to present an orthogonal data input set containing the largest amount of information possible. This gave some encouraging results, but was neither sensitive nor accurate enough for clinical use. A set of experiments was performed to give local information on small windows of scintigram images. By this method areas of abnormality could be sent into a subsequent classification network to diagnose the cause of the defect. This automatic method of detecting potential defects did not work, though the networks explored were found to act as smoothing filters and edge detectors. Network design was investigated using genetic algorithms (GAs). The networks evolved had low connectivity but reduced error and faster convergence than fully connected networks. Subsequent simulations showed that randomly partially connected networks performed as well as GA designed ones. Dynamic parameter tuning was explored in an attempt to produce faster convergence, but the previous good results of other workers could not be replicated. Classification of scintigrams using manually delineated regions of interest was explored as inputs to ANNs, both in raw state and as principal components (PCs). Neither representation was shown to be effective on test data.