Logo of the Physikalisch-Technische Bundesanstalt

Using machine learning for quality assurance in mammography


In mammography screening, the risks posed by the X-rays used must be carefully weighed against the potential gain of a timely breast cancer diagnosis. e rule of thumb: the higher the applied radiation dose, the better the resulting image quality and, therefore, the greater the likelihood of reliably recognizing breast cancer at an early stage. For this reason, an objective measure of image quality, which can be determined as accurately as possible, is needed in order to achieve an optimal compromise for the applied radiation dose.

Using mammography, suspicious structures in breasts can be detected and treated early on. (Photo: dpa)

Nowadays, image quality in mammography is determined based on the degree to which fine structures can be detected in a technical phantom. Whereas in the past, the images were examined primarily by radiologists who – successfully (or unsuccessfully) – detected the signals, mathematical procedures (socalled “model observers”) are used today. Several errorprone data processing steps are necessary for this in order to reliably determine the image quality from a number of images.

PTB has developed an alternative procedure using modern machine learning methods. With this procedure, the image quality of mammographs can, for the first time, be determined automatically by means of individual images. The new method is robust and significantly more precise than the procedures used up to now.