Fast automatic myopic deconvolution of angiogram sequence

We present a fast unsupervised myopic deconvolution method dedicated to quasi-real time processing of video sequences such as angiograms. Our method is based on a Bayesian approach of which the tuning parameters are automatically set thanks to the marginalized likelihood of the observed image. We demonstrate the effectiveness of our approach on simulated and empirical images. 1. CONTEXT AND PREVIOUS WORK Coronary angiography is a medical imaging technique used to visualize heart vessels (the coronaries) in order to diagnose and prevent potential heart failures. Its principle consists in injecting a radiocontrast agent by catheter into an artery and capturing an image sequence of the cardiovascular system thanks to X-ray irradiation. In a previous work [1], we shown that the blur of angiogram sequences could be approximated by a convolution by a shift-invariant point spread function (PSF). We found that the actual PSF is constant for a given video sequence but depends on the operating conditions and on the patient. We therefore proposed to use multi-frame blind deconvolution to improve the quality of the videos both in terms of signal to noise ratio and resolution. We observed that the PSF are approximately isotropic and bell shaped with a profile similar to a Lorentzian distribution: h(r) ≈ η 1 + ‖2 r/γ‖ , (1) with r the 2-D position on the detector, γ the full width at halfmaximum (FWHM) of the PSF and η a normalization factor. This multi-frame blind deconvolution method is however lengthy and requires the tuning of a number of control parameters which make it impracticable for non-specialists and incompatible with the quasi real-time requirement. In this paper, we propose several improvements to achieve a fast and unsupervised method which works well in practice. 2. UNSUPERVISED MYOPIC DECONVOLUTION 2.1. Data model For a given frame of the sequence, a raw image writes [1]: