Sequential Analysis of Medical Images Using Neural Networks
Marek Czeleń, Grzegorz Wojtas, Zbigniew Mikrut · 2003
A problem frequently encountered in the analysis of medical images is reaching a correct and reliable diagnosis of the examined organ, particularly by persons with little experience, e.g. first year medical students. That fact has become an inspiration for construction of an assisting program, based on application of three neural networks: 1) moving the attention point along the selected object, 2) elaborating the diagnosis for the area in the current viewport, 3) reporting the end of the examination. In the present work a summary is done for almost a hundred of experiments, testing the efficiency of all those neural networks in the analysis of raw medical images. The final version of the algorithm uses backpropagation type networks: it operates on square areas which are 31×31 pixels in size for the first and third networks and 91×91 pixels in size for the second (diagnosing) network. The achieved level of correct recognition (about 80%) is a very good result, if one takes into account rather low quality (dynamic range) of the analysed images and the observed natural interferences. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.