Deep learning is a subfield of artificial intelligence that uses a cascade of simple nonlinear functions for feature extraction and decision-making. We apply this technique for automated quantification of diagnostic PET and CT images. Software packages for analysis of diagnostic images that are based on our previous research are currently used at approximately 1,000 hospitals worldwide.


Reading and reporting diagnostic images is a subjective process. Even though the physicians involved in this process are specialists in radiology or nuclear medicine, and thereby meet the basic standards, several studies have shown that image interpretations from different physicians vary substantially. One approach to reduce variability among physicians and to control the quality is to make the reading and reporting process more objective by the use of quantification of images, for example calculation of tumor burden to the skeleton in prostate cancer patients or left ventricular volumes and ejection fraction in patients with coronary artery disease.

Our contributions in this field have been innovative in that we have developed fully automated methods eliminating the need for operator manipulation. With our approach the quantitative results will be the same in all hospitals performing an examination.