[フェロー記念講演] Image Analysis and Medical Applications Forty Years

This special talk is set to commemorate the author’s fellow title of IEICE Japan. Researches between these forty years related to image analysis and medical applications is introduced in this talk. They include wide range of research of learning, registration, classification method, spatio-temporal image processing, 3D display, and sign language communication-aid system. Keyword Medical image,3D display, Registration,Sign language,Fuzzy relation It was 1966 when the author begun the research of pattern analysis/recognition at Graduate School of Engineering Science, Osaka University. This special talk is set to commemorate the author’s fellow title of IEICE Japan. Researches between these forty years related to image analysis and medical applications are introduced. 1. Basic research for pattern analysis 1.1. Learning of unknown pattern The author’s first major research work as a graduate student was learning and recognition of unknown signal pattern. When the pattern shape is unknown under the Gaussian noise environment but the pattern class is given by a teacher, pattern classification is a simple optimum estimation problem (learning with teacher). However, when the pattern class is unknown, it is a nonlinear estimation problem. One solution to such problem is to compensate the instruction of the teacher by a decision of the classifier itself (decision-directed machine). Another approach is to exclude outliers, and this improves the estimation accuracy (robust estimation). However, these approaches give biased estimator. To obtain the unbiased estimator, we proposed some machines including a non-decision-directed machine for two class problem [1]. Further, some other problems including time-varying pattern [2], synchronization for unknown period of sequence [3], and with feedback link [4] were studied. 1.2. Classification method based on fuzzy relation Besides the above, the author dealt with fuzzy set theory, and developed fuzzy classification method based on fuzzy relation [5]. Although Prof. Zadeh, the originator of fuzzy theory seems also felt need of such method, I could publish the paper earlier than him. This paper was introduced in many books, and after 21 years later from the original publication, reproduced paper was published from IEEE Press [6] as well as the extended analysis paper [7]. An extension of the theory is awarded with the Best Paper Award in Japanese Society for Artificial Intelligence Annual Convention. 1.3. Registration: Esp. for industrial applications Image registration is one of the most fundamental techniques in image analysis. As a matter of course, it is important in medical image analysis and also in industrial applications such as dicing saw for silicon wafer. Functionally it is also called "positioning", "localization", or "alignment". We developed an alignment system for IC wafer with Disco Co. which is the leading company in the world in dicing saw. First we developed an alignment system using straight line detection [8], which uses direct detection of straight lines on surface of IC wafer. Direct straight line detection method using sequential detection algorithm is faster than Hough transform [awarded with Daily Industrial Newspaper Ten Most Important New Product Award]. Then, we also developed a pattern matching circuit based on normalized correlation for practical use in dicing saw in 1984 [9]. In that time it was impossible to calculate normalized correlation in real time. So, we developed a normalized correlation between only on pseudo-random sampling points selected one point from each raw and each column. This type of hardware may be the first one in the world though Cognex, leading industry in machine vision spinned out from MIT, declares in their home page [10] they are the first in 1987. However, it should be noted that Cognex developed a pattern matching machine for industrial applications using mutual information soon after Viola (1995) 1.4. Spatio-temporal image: Hough, Radon, and LG transform / filtering for tracking blood cells Motion of blood cells (platelet stained by fluorescent dye (Fig.1), leukocyte, etc.) is depicted as lines of traces in 2D spatio-temporal images [11]. Then, they can be detected by directional Laplacian-Gaussian filter [12]. On the other hand, if they are straight line, they can also be detected by Hough transform. We have shown that Laplacian-Gaussian filter and Hough transform are essentially the same detection method, and also the same as Radon transform [13](Fig.2). Fig.1. Enlarged image sequence of Fig.2. Spatio-temporal moving fluorescent dot (see arrow) in image along capillary scanning laser ophthalmoscope (SLO). vessel (a), and extracted trace by LG filter (b). 2. 3D display: Esp. for echocardiography 3D display is the most basic function of 3D image processing. 2.1 3D image processing On-line connection of the echocardiograph to the computer was not so popular in the 1970s. Then, we scanned the tomograms by a FSS (flying spot scanner) connected to PDP-8 computer, which are taken at intervals of 5mm using 2-3 MHz ultrasound with cardiac synchronization. Then we extracted boundary lines of each tomogram and composed a binocular stereoscopic image on a Fig.3. Specifying desired plane (upper) and extracted ASD (lower). storage type CRT [14,15]. Also they are processed to extract arbitrary plane image [14,16] (Fig.3) and measured volume of