Application of neural networks to lesion detection in SPECT

An artificial, single-layer, feedforward neural network was developed for lesion detection in single photon emission computed tomographic (SPECT) images. The network has 121 input nodes and one output node. A backpropagation algorithm with a sigmoid activation function is used for its supervised learning. The diagnostic performance of the neural network is studied at various noise levels for different numbers of training images using receiver operating characteristics analysis. Three noise levels (30000, 50000, and 100000 counts/slice) were used. At each noise level two training data sets (one with 40 and one with 200 simulated SPECT images) and a testing data set (200 SPECT images unknown to the trained network) were generated. The diagnostic task was the detection of a cold lesion 1.0 cm in radius at a known location of the image. The neural network trained quickly and accurately in all cases. Network convergence to a minimum in cumulative squared error was easy to achieve regardless of the noise level or the number of training images. However, the performance of the network as a decision maker for lesion detection on new images was very much affected by the number of training images. When trained with a sufficiently large number of images the performance of the network was significantly improved, particularly in the low count (high noise) case.<>