Skinfold Recognition for Positioning Quality Evaluation on Mammography

Skinfold is one of the positioning issues in mammography quality evaluation. As it overlays breast tissues in mammograms, it implies that splints over-compresses the breast or the patient moves slightly. Skinfold affects radiologists in screening mammography. Therefore, radiographers adjust their operations during taking mammograms. In this paper, we propose a skinfold recognition method based on a convolutional neural network and extract features for analyzing the skinfold recognizer. The dataset consists of patches annotated with positive-class or negative-class, which is the consistent decision-making by four radiographers. Our method performs DenseNet-121 to train a skinfold recognizer and compares Vgg-16 and ResNet-101. Furthermore, extracting the class activation map is for interpreting what features the skinfold recognizer detects. The experiment results demonstrate that the proposed scheme outperforms the compared approaches.