This paper discusses the problem of the feature extraction from the standing position-anteroposterior radiogram on the statistical basis as a part of the study for the purpose of computerizing the mass screening of stomachs with a particular interest in the deformity of the apex, or the sulcus angularis.
As to a number of medically confirmed samples, features were evaluated on their performance to discriminate among the normal samples, ulcerated samples and cancerous samples by a typical method of the discriminant analysis, the minimum Mahalanobis' distance classification. The error rate was estimated by the leaving-one-out method and then the best sets of discriminative features were selected on the basis of the forward sequential procedure. The separability of the samples and the characteristics of the selected features were also inspected by the cluster analysis.
Two approaches were examined in the feature extraction. One consisted of extraction from the contour of the stomach. The best sets of features were selected from among five geometrical measurements and twenty-eight Fourier descriptors. In the best N/A, or normal/abnormal, discrimination in the discriminant analysis, the estimated error rate was 7.5 percent ; in the N/U/C, or normal/ulcer/cancer, discrimination, it was 25 percent. These figures agreed with the results of the cluster analysis and the inspection of the contour by six physicians.
The other approach was extraction of the features through the two-dimensional orthogonal transform of the grey level image of the region of the apex. The selected best set of features for the N/A discrimination resulted in the error rate of 6.4 percent; the best set for the N/U/C discrimination resulted in that of 21.3 percent. These figures also agreed with the results of the cluster analysis and the inspection of the images by an experienced physician.
In both approaches, the selected features tended to belong to the different clusters of the features. It is clearly recognized in the N/A discrimination and supports the validity of the selected features.
The following conclusion may be drawn from the experiments :
1. It has been proved that there exist feature sets of the error rate comparable to that of physicians.
2. While discrimination between normal and abnormal samples was good, that between ulcerated and cancerous samples, on the contrary, was poor.
