Preprocessing of biomedical images for neurocomputer analysis

Thermal infrared images were analyzed to evaluate the growth of tumors in laboratory mice. Special segmentation algorithms were used to extract the areas of interest. The data were then normalized and used to train a three-level backpropagation neural network which was able to learn to distinguish between seven stages in a tumor growth. Following this, the data were level-sliced and feature attributes were extracted in the spatial frequency domain. After all preprocessing, a much smaller and faster backpropagation network was able to accurately classify the seven different tumor stages from less than 1% of the original image data. Classification accuracy was maintained, even when up to 30% noise was added to the images.<>