• New methods of artifact reduction of the images acquired by magnetic resonance imaging
• On the optimum architecture of the biologically inspired Hierarchical Temporal Memory model

Result 1: New methods of artifact reduction of the images acquired by magnetic resonance imaging

Project: VEGA 2/0142/08

Authors: V. Juráš,  L. Valkovič,  P. Szomolányi, I. Frollo, L. Vojtíšek, T. Dermek

New methods for enhancing the quality of the images from MRI (Magnetic Resonance Imaging) were designed, experimentally tested and clinically validated. The focus has been put on the images biased by the noise, geometrical distortions, moving and stress artifacts, and mostly on eliminating of magnetic field in-homogeneities causing the geometrical distortions.

For defined ROIs (Region-of-Interest) in various biological tissue (cartilage, inter-vertebral disc, tendon), the correction algorithm for relaxation constant estimation has been designed. The stress effect has been tested by comparison of the control extremity (hand) and the extremity under the stress.

Proposed methods have been successfully tested in the patients after cartilage transplantation, in the brain examination using fMRI, and in the stress evaluation. The method has been applied in the patients with MACT (matrix associated chondrocyte transplantation) for better distinguishing between native and transplant cartilage. It is very useful also in post-operative monitoring of such patients.

International cooperation: Univ.-Prof. Dr. Siegfried Trattnig, MR Center, Highfield MR, Department of Radiology, Medical University of Vienna, Austria. Agreement for Scientific Cooperation, 28.8.2006 and 24.8.2009.

Figures: A) The sagittal map of human knee in-vivo corrected with coefficient of determination stored in the process of the map calculation. B) The comparison of the human brain images a) morphological image, b) the image after noise correction.

Publications:

• JURÁŠ, Vladimír - ZBÝŇ, Š. - SZOMOLÁNYI, Pavol - TRATTNIG, S. Regression error estimation significantly improves the region-of-interest statistics of noisy MR images. In Medical Physics, 2010, vol. 37, no. 6, p. 2813-2821. ISSN 0094-2405. (2.704 - IF2009).
• VALKOVIČ, Ladislav - JURÁŠ, Vladimír - DERMEK, Tomáš - VOJTÍŠEK, Lubomír - FROLLO, Ivan. The effect of stress on MR image contrast in the human hand at low field NMR. In ELITECH '10 : 12th Conference of Doctoral Students. - Bratislava : Faculty of Electrical Engineering and Information Technology, Slovak University of Technology, 2010. CD-ROM, p.1-7.
• JURÁŠ, Vladimír - ZBYN, S. - SZOMOLÁNYI, Pavol - FROLLO, Ivan - TRATTNIG, S. The improvement of region-of-interest statistics in musculoskeletal MRI. In ISMRM-ESMRMB 2010 : Joint Annual Meeting. Stockholm, Sweden, May 1-7, 2010.
• VALKOVIČ, Ladislav, - WINDISCHBERGER, C. Method for geometric distortion correction in fMRI based on three echo planar phase images. MEASUREMENT SCIENCE REVIEW, Volume 10, No. 4, 2010, p. 116-119. ISSN 1335 - 8871.   

Result 2: On the optimum architecture of the biologically inspired Hierarchical Temporal Memory model

Projects: VEGA 2/0019/10

Authors: S. Štolc, I. Bajla 

In recent years of neuroscience research, the attention has been focused on various biologically inspired models of representation and processing of visual and other kinds of sensory data. The Hierarchical Temporal Memory (HTM) is one of such models. Our research has been concentrated on methods of optimization of important controlling parameters of the HTM model, applied to the task of visual pattern classification. We have proposed a new methodology for constructing the optimal HTM network architecture, which guarantees a balanced utilization of all input pixels. The attention has been also paid to new, more application-specific, methods for generating pattern sequences required for successful training of the HTM network. Another original contribution has been achieved by development of algorithms for optimal setup of the vector quantization which is based on the "box counting" method borrowed from the chaos theory. For optimizing the inference parameters within each HTM node, a specialized method, maximizing the entropy of belief distribution, has been proposed. The obtained classification accuracy of the single-layer HTM network, optimized by the proposed methods, has been compared to various known classification methods and testified perspectives of the entire HTM approach. A very good rank , within a class of the state-of-the-art classification methods, of the actual optimized HTM model has been achieved. Besides this result, we consider the optimization algorithms for various parameters of the HTM model themselves for valuable contribution to further development of this very promising biologically inspired models in the field of artificial intelligence.

Cooperation with universities: Department of Applied informatics, Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava (Assoc.Prof. I. Farkaš, PhD.)
Foreign partner: AIT, Austrian Institute of Technology GmbH, Seibersdorf, Austria

Figure: Example of a 2-layer HTM network exposed to an image input of size 20x20 pixels.

Publications: