| This thesis shows that competitive neural networks can extract general features from digital photomicroscope images of four types of human breast disease (fibrocystic change, fibroadenoma, infiltrating duct carcinoma and infiltrating lobular carcinoma). The competitive neural networks employed in this thesis include (i) the basic competitive neural network with conscience, (ii) one-, two- and three-dimensional self-organizing feature maps, and (iii) learned vector quantitizer. The networks were trained with training sets taken from digital images of representative areas of breast disease in microscopic sections. The vectors varied from four to as many as 10,000 units in length. The pixel blocks were always square, containing from 2 x 2 up to 100 x 100 pixels. The number of features extracted were as few as four to as many as 72. The training time varied from as little as 50 epochs to as many as 2,000,000 (approximately two months time on the PC). Each network was compared to the others for performance, as well as to the performance of a pathologist. (Abstract shortened by UMI.)... |
