Introduction: Detection of various artifacts in images, caused by many factors, is essential for evaluating quality control of mammography units that is daily performed by the radiographer. The paper addresses all the important issues in the process of building an automatic system for artifact detection: image acquisition and labeling, image preprocessing for feature extraction, feature extraction and selection, and classification modeling with evaluation. Purpose: The aim of this research was to build a system for the automatic detection of artifacts in mammographic images that can be used in practice as an analysis tool for the daily quality control of flat-field homogenous images in mammography systems. Methods: We defined the procedures for image acquisition, developed measures for feature extraction and tested different classification methods for artifact detection. Our system was designed to recognize artifacts on the basis of contrast non-uniformity, ghosting, lines, dead pixels and acquisition errors, which are the most frequently observed artifacts in mammographic and other DR radiographic images. Results: We acquired 542 mammographic images produced with 7 different mammography units, of which 101 images have non-uniform contrast, 53 have ghosting artifacts, 52 have line artifacts, 18 have dead pixels and 51 have artifacts due to acquisition errors. The artifact detection was performed using different features and different modeling techniques. We designed nine different features, which measure the uniformity, texture information, average gray-level segment difference and the presence of lines and dead pixels. Different combinations of features were tested with five classification methods: logistic regression, na?ve Bayes, decision tree, random forest and neural network with a multilayer perceptron. A leave-one-out cross validation yielded the best overall results (measured by the AUC in the ROC analysis) of 98.6% for detecting non-uniform contrast, 92.2% for detecting ghosting artifacts, 95.1% for line artifacts, 100% for detecting images with dead pixels and 97.8% for detecting artifacts due to acquisition errors. Discussion and conclusion: The best overall results were achieved by using the random forest for classification together with all nine features or in some cases just with the specially designed features for the specific artifact.
|