We proposed a method for a computer-aided diagnosis system that distinguishes between benign and malignant lesions in gastrointestinal digital radiography. To begin with, the level set method was applied in order to extract a tumor region from the image which was smoothed by the bilateral filter. Next, we selected four image features with the large SN ratio among various image features obtained from a tumor region using the Mahalanobis-Taguchi method, which has been employed mainly in quality engineering. The selected four image features—circularity, irregularity, size, and perimeter—were used as input data for the artificial neural network, which was employed for distinction between benign and malignant lesions. By using 43 regions of interest cropped from the 43 clinical cases, the area under the ROC curve (AUC) of diagnostic accuracy for the classification obtained with this proposed method was 0.970, whereas the average AUC obtained with 7 human observers (3 radiologists and 4 radiological technologist) was 0.941.