Blind Deconvolution for Confocal Laser Scanning Microscopy

Confocal laser scanning microscopy is a powerful technique for studying biological specimens in three dimensions (3D) by optical sectioning. It permits to visualize images of live specimens non-invasively with a resolution of few hundred nanometers. Although ubiquitous, there are uncertainties in the observation process. As the system's impulse response, or point-spread function (PSF), is dependent on both the specimen and imaging conditions, it should be estimated from the observed images in addition to the specimen. This problem is ill-posed, under-determined. To obtain a solution, it is necessary to insert some knowledge in the form of a priori and adopt a Bayesian approach. The state of the art deconvolution and blind deconvolution algorithms are reviewed within a Bayesian framework. In the first part, we recognize that the diffraction-limited nature of the objective lens and the intrinsic noise are the primary distortions that affect specimen images. An alternative minimization (AM) approach restores the lost frequencies beyond the diffraction limit by using total variation regularization on the object, and a spatial constraint on the PSF. Additionally, some methods are proposed to ensure positivity of estimated intensities, to conserve the object's flux, and to well handle the regularization parameter. When imaging thick specimens, the phase of the pupil function due to spherical aberration (SA) cannot be ignored. It is shown to be dependent on the refractive index mismatch between the object and the objective immersion medium, and the depth under the cover slip. The imaging parameters and the object's original intensity distribution are recovered by modifying the AM algorithm. Due to the incoherent nature of the light in fluorescence microscopy, it is possible to retrieve the phase from the observed intensities by using a model derived from geometrical optics. This was verified on the simulated data. This method could also be extended to restore specimens affected by SA. As the PSF is space varying, a piecewise convolution model is proposed, and the PSF approximated so that, apart from the specimen, it is sufficient to estimated only one free parameter.