Research project

Medical image reconstruction is an important tool for clinical diagnostics in order to provide 2D and 3D views of patients' inner organs. The image quality of the reconstructed volumes depends (often heavily) on both modifiable and intrinsic parameters of the imaging system and process. Additionally, image quality itself has to be understood in terms of image properties that are required for diagnosis which again is influenced by the personal sensation of the human observer.

In digital breast tomosynthesis (DBT) data acquisition is subject to several restrictions (limited angle, low dose) which makes optimal setting of the reconstruction parameters mandatory to achieve competitive image qualities. Mammographic images mainly serve as an instrument for early detection of breast cancer - the top cancer in women according to WHO (2014) [1]. Therefore the demand of optimal image quality in DBT necessitates bringing in line preservation and best possible detectability of lesions like microcalcifications, masses or spiculations with reduction of noise induced by the imaging system.

The goal of this project is to provide tools that can deal with multidimensional parameterspaces to estimate optimal reconstruction settings with respect to pre-defined observer requirements. Combination of techniques from machine learning with a parameterized reconstruction algorithm will be used to improve the diagnostic value of tomosynthesis images.

[1] World Cancer Report 2014, IARC, Lyon 2014.