In order to reliably diagnose breast cancer from mammography images, the small contrast difference between cancerous and healthy tissue needs to be resolved. Regular quality assurance check-ups are carried out that assess the image quality of mammography devices. For this, multiple images of a technical phantom are recorded and subsequently analyzed. Image quality is then expressed as the contrast-detail curve which summarizes the detectability of various contrast differences in the image of the technical phantom. The contrast-detail curve is determined by applying cumbersome pre-processing and regression procedures. Multiple images are required, and a European guideline recommends taking at least 16 images in order to ensure a reliable judgement about the image quality.
In this work deep learning is applied to provide an alternative path for determining the contrast-detail curve. A convolutional neural network (CNN) is trained on a large database of mammography images. The trained network predicts contrast-detail curves from single images without any pre-processing. A similar accuracy is reached as for the current procedure that requires 16 images. In this way, the workload for routine quality assurance measurements can be significantly reduced. It is shown using explainable AI that the trained network focuses particularly on those regions in the image that contain the relevant information about the contrast-detail curve. Our results demonstrate that mammography image quality assessment can benefit from deep learning, and additionally they provide insight into how deep learning performs this task of image regression.