Logo der Physikalisch-Technischen Bundesanstalt

A new explainability approach for deep learning applied to image quality assessment in mammography


Image quality assessment is of particular relevance in image processing applications. This is especially true in mammography, where it helps to achieve a high detection quality at the lowest possible radiation dose. The assessment of image quality in mammography is carried out in accordance with the recommendation of a European Guideline. Recent research has shown that the use of deep neural networks can improve this established procedure. However, neural networks generally suffer from a lack of interpretability and therefore their behavior needs to be studied more thoroughly prior to recommending their routing use.


We developed a new explainability approach – oriented, modified integrated gradients (OMIG) and applied it to the recently developed deep learning approach for mammography image quality assessment. As a result, OMIG provides a more meaningful interpretation of the predictions made than other established explainability methods. Furthermore, it confirms the validity of the developed deep learning approach for mammography image quality assessment and increases the trustworthiness into its application.


The research is published in Machine Learning: Science and Technology Journal and will be presented at the Medical Image Understanding and Analysis (MIUA 2022) Conference and at the 53rd Annual Conference of the DGMP (German Society for Medical Physics).


The paper is available at Opens external link in new windowhttps://doi.org/10.1088/2632-2153/ac7a03


Contact: Narbota Amanova, 8.42