Performance analysis of data mining and deep learning methods in diagnosis with medical image processing

Image processing is a set of methods in which different mathematical algorithms are applied in order to obtain useful information from images transferred to the computer. In this method, the primary goal is to improve the quality of images. Another positive aspect of the image processing method is the acquisition of numerical values (attributes) that describe the images. The features obtained from the images by the image processing method are used to classify the images. Although imaging methods are advanced, statistical methods and various image processing algorithms may also be important in diagnosing various diseases early. In this study, images were digitized using medical image processing techniques and its success in detecting normal-abnormal lesions and then in distinguishing malign-benign lesions was examined. The effects of five different image pre-processing algorithms, two different segmentation methods and breast tissue type on the results were investigated. In addition, classification performances of data mining and deep learning algorithms in different image processing techniques used were compared. For this, 322 mammography images, 209 normal, 61 benign and 52 malign, in the open source MIAS database were used. When five different pre-processing and two different segmentation algorithms created in image processing were compared, it was seen that the preprocessing and image segmentation techniques applied on images actually affected the classification performance measures. In general, SVM, RF and ANN, which are classical data mining methods, performed better than other algorithms in this data set. Deep learning methods also gave similar results to classical data mining methods.