Dynamic relaxation in algebraic reconstruction technique (ART) for breast tomosynthesis imaging

A major challenge in Digital Breast Tomosynthesis (DBT) is handling image noise since the 3D reconstructed images are obtained from low dose projections and limited angular range. The use of the iterative reconstruction algorithm Algebraic Reconstruction Technique (ART) in clinical context depends on two key factors: the number of iterations needed (time consuming) and the image noise after iterations. Both factors depend highly on a relaxation coefficient (λ), which may give rise to slow or noisy reconstructions, when a single λ value is considered for the entire iterative process. The aim of this work is to present a new implementation for the ART that takes into account a dynamic mode to calculate λ in DBT image reconstruction.

A set of initial reconstructions of real phantom data was done using constant λ values. The results were used to choose, for each iteration, the suitable λ value, taking into account the image noise level and the convergence speed. A methodology to optimize λ automatically during the image reconstruction was proposed.

Results showed we can dynamically choose λ values in such a way that the time needed to reconstruct the images can be significantly reduced (up to 70%) while achieving similar image quality. These results were confirmed with one clinical dataset.

With simple methodology we were able to dynamically choose λ in DBT image reconstruction with ART, allowing a shorter image reconstruction time without increasing image noise.