演題番号 : P1-r18
大羽 成征 / Shigeyuki Oba:1 石井 信 / Shin Ishii:1
1:京都大学 情報学研究科 / Graduate School of Informatics, Kyoto University 2:日本科学技術振興機構 / Japan Science and Technology Agency
Confocal microscope has been broadly used to observe complicated three-dimensional structure of single or multiple neurons. Since sub-cellular structures of neurons such as spines and synapse binding cites are difficult to find out by the present technology of confocal microscopy, there have been many efforts to develop image processing tools based on statistics. In this work, we developed a new pre-processing method based on diffusion tensor field estimation. Image processing with diffusion tensor field estimation is an extension of a well-established technique of image deconvolution based on point spread function (PSF). The PSF deconvolution is often used in pre-processing phase to reduce noise and achieve higher resolution by gathering pixel information that has been diffused into nearby pixels. However, appropriate radius of PSF to effectively reduce noise may be different between images with different depth in confocal image stacks, or even in different regions in a single image. Appropriate shape of the PSF is not common between images. To deal with the problem above, we considered PSF with an arbitrary tensor parameter of diffusion that is a function of coordinate in the image stack. We also estimated the diffusion tensor field that is smooth over a single image. In addition, we developed a methodology to estimate a hierarchical diffusion tensor field from a given image. When applied to several confocal microscopic image stack datasets, our new method exhibited improvement in detectability of sub-cellular structures of neurons.