[The Development of Algorithm of Intellectual System of Supporting Decision-Making in Mammographic Diagnostics of...

The article considers issues of training models of convolutional neuronic network (CNN) for automated identification of point functions of visualization to discern mammography pictures belonging to negative, false benign and malignant cases, targeting to improve interpretation of results of mammographic examination. On the basis of automated deep training (application of CNN) the approach is proposed, that permits to detect small differences in mammographic pictures to diagnose true positive and false positive results. The information base for training CNN is represented by digital base (open resource for methods of intellectual studying of analysis of mammographic pictures) of data of mammography screening of the Massachusetts General Hospital and the Winston-Salem Medical College, the USA. The share of false positive results according trained model in the total number of examined patients from different age groups made up from 22% to 32%. To compare, share of false positive results based on decoding of mammographic picture by medical specialist varied from 34% to 53%. The trained model based on the CNN can be used for mammographic pictures of any database. The predictive accuracy of model depends, among other things, on volume of training sample. Therefore, formation of open database of results of mammographic examinations with accurately established diagnosis will permit later on to broadly apply possibilities of deep training in medical practice. The study has great potential to incorporate deep training of CNN into clinical practice of screening of breast cancer and to improve interpretation of mammographic pictures.