A boundary 1 of a wall in an initial image of time-series images is decided automatically or manually. A boundary 2 is generated automatically or manually in a neighbor of the other boundary of the wall. A normalized image is generated on the basis of an image pattern of a region surrounded by the boundaries 1 and 2 and registered as a template image. In each of the time-series images, boundaries 1 and 2 that generate a normalized image most similar to the template image are calculated. In this manner, wall thickness change can be calculated automatically stably and accurately even in the case where luminance change in the boundaries of a wall is obscure. This method is useful for diagnosis of heart disease etc. in medical image diagnosis.