Department of Image Processing

Volume reconstruction (Ing. Martin Čapek, Ph.D.)

A system of methods leading to volume reconstruction of biological specimens larger than the field of view of a confocal laser scanning microscope (CLSM) is presented. In the first step large tissue specimens are cut into thin physical slices. Second, volume image data sets (spatial tiles which overlap) are captured from all studied physical slices by CLSM. The third step is merging of overlapping spatial tiles of the same physical slice in horizontal direction (mosaicking) using a registration algorithm based on rigid-body geometrical transformation. In the fourth step volumes of successive physical slices are merged in axial direction by applying an elastic registration algorithm that gives us possibility to compensate for deformations of objects due to the cutting of the specimen. We also describe a method evaluating errors of elastic registration using a stereological measurement (Point Grid), and helping to keep true objects morphology, since elastic registration does not maintain correctly objects sizes and shapes when they change in space (enlargement or diminishing). Finally, the images are enhanced to compensate for the effect of the light attenuation with depth occurring in confocal microscopy. The resulting large digital volumes are visualized by a hardware accelerated volume rendering. The presented methods are demonstrated by a reconstruction of the middle part of a 21-day-old laboratory Norway rat embryo.

Elastic alignment (Ing. Jan Michálek, Ph.D.)

We apply elastic registration in the framework of volume reconstruction, where an object acquired by CLSM from parallel physical sections is composed and mutual positions of the sections including deformations caused by their cutting have to be found. Our aim was to find a parallelizable algorithm that can be implemented on a graphics card using NVidia CUDA programming environment.
The correspondence between two images to be registered can be found by minimization of a functional penalizing the dissimilarity of corresponding image elements together with roughness of the correspondence function. The functional consists of two parts where the first part is the discrete total variation as a measure of roughness and the second one represents L1 norm as a measure of dissimilarity of images. The proposed functional is well-suited to be solved by optimization of (max,+)-labelling problems. A parallelizable version of these optimizations represents an equivalent transformation of a (max,+)-labelling problem. Then the functional can be computed in a parallel way using horizontal and vertical lines of images only.
The proposed elastic registration algorithm was implemented both running on CPU using Matlab and C language and running on a graphics card using Matlab and NVidia CUDA programming environment. We found that CUDA-based implementation of the algorithm is approx. ten times faster than CPU-based implementation, depending on the size of images to be registered.

Image analysis & Total variation (RNDr. Jiří Janáček, Ph.D.)

Various tasks of image analysis and computer vision can be solved by optimization of suitable cost functions. An example is image filtration with cost function penalizing image roughness and distance from the original image, where total variation (TV) and Lp distance can be used. Another example is regularization of geometrical models of objects captured on grayscale images, for models of fibre-like structures total length can be used for penalization of rougness and negative sum of image values in nodes can be used as data term of cost function. The minima can be found by steepest gradient descent using minimal cut in a suitable graph in each iteration.
Examples of TV-L1 filtration of TEM images of nanoparticles and smoothing of the model of capillaries on 3D confocal image will be shown.