TU-A-218-06: Dose Reduction by Iterative Reconstruction of Planar or Tomographic X-Ray Images

Abstract

Purpose: Significant dose reductions in X‐ray examinations such as planar imaging,tomosynthesis, and CT can be achieved by mathematically modeling the physics and noiseproperties of each element of the image acquisition chain. Methods: Low‐dose X‐ray images suffer from high quantum and electronic noise. In order to reduce the effects of this noise, an estimate of the ideal underlying noise free image is reconstructed based on an accurate model of the imaging physics. This model takes into account the properties of the X‐ray tube, scintillator crystal, light sensitive TFT matrix, and readout electronics. The calculation of the estimate is based on compressive sensing theory which was extended to Poissonnoise. Compressive sensing allows for accurate reconstructions of three dimensional images based on only a small number of projections by exploiting very general image compressibility properties in a suitable representation. By a similar principle, planar images can also be reconstructed. A natural choice of representation are two and three dimensional wavelets but other, more problem specific bases are possible. An iterative reconstruction algorithm based on these results was implemented on a GPU for fast processing. Results: The modeling of the image acquisition chain and the extension of compressive sensing theory to Poissonnoise yields an iterative algorithm that is able to deliver high quality two and three dimensional reconstructions from a small number of noisy projections. Conclusions: The work shows that large dose reductions in X‐ ray examinations (between 50% for planar imaging and 90 –95% for CT) are possible by saving on both the number of projections and the dose per projection. The former results in reduced spatial information and the latter introduces more noise into each projection. Both effects are compensated for by an accurate physical model and by exploiting natural image sparsity information.