In this paper, a procedure for recognizing blood vessel shadows (VSL) in the lung in the chest photofluorograms is presented together with experimental results.
The construction of the presented procedure is basically the same as that which was used in the software $ystem for automated interpretation of chest roentgenograms AISCR-V 2 developed previously by the authors' group. This procedure consists of three steps; (1) extraction of candidate components, (2) calculation of features and (3) classification. In the first step, the second order differential linear filter is applied to the digitized input picture. Connected components are extracted from the filter output by the threshold operation. In the second step, five features about the position, the area, and the average density are measured from each connected' component. Components, of which the feature values are far from those admissible as VSL's, are eliminated. The remaining components are considered as candidates for VSL s. In the last step, candidates of VSL's are classified into VSL and others using linear discriminant functions which were designed by an iterative learning procedure based upon the training sample set. The procedure was applied to thirty-seven chest photofluorograms taken in mass screening, and all of main parts and about, 82% of smaller, parts of vessel images were recognized correctly. By adding this procedure to the previous system AISCR-V 2, it becomes possible for the first time to construct the automated pattern recognition system of chest, photofluorograms which makes it possible to recognize all of three major component patterns-heart, ribs and vessels.
抄録全体を表示