The computer analysis and diagnosis of gastric ulcers.

Abstract

THE VALUE of roentgenographic examination of the stomach in patients with upper gastrointestinal symptoms is well established (10, 12), and the accuracy of the roentgenologist in detecting all stomach abnormalities has been variously cited to be between 85 and 95 per cent (1, 3, 4, 7, 10). Even so, one of the most controversial differential diagnoses confronting a radiologist is that of whether a gastric ulcer is benign or malignant (15). Many roentgen signs have been described which are thought to be of value in making this differentiation, but the statistical significance of the various clinical findings and roentgenographic signs has never been established. The use of computers in analyzing data relating to medical disease entities is widely accepted. Computer diagnosis has exceeded the diagnostic accuracy of the trained radiologist in selected bone tumors (8). Radiographic manifestations of carcinoma of the lung have also been evaluated by computer technics (9). The statistical evaluation of 70 radiographic signs and clinical findings pertaining to benign and malignant gastric ulcers is presented in this communication. Computer versions of standard statistical tests have been utilized. From the statistically significant information so obtained, a computer program for diagnosing malignant versus benign gastric ulcers has been developed and tested. Materials and Methods In the review of 65,000 hospital admissions, 496 histologically proved benign or malignant gastric lesions were encountered. Of these, 178 were incidental findings at autopsy or at surgery without previous x-ray examination. Of the remaining 318 patients, 93 had roentgenograms suitable for retrospective analysis. There were 68 benign and 25 malignant ulcers. Forty-two cases of nonulcerating gastric neoplasms were also found. Following a review of the literature, a list of 70 clinical and laboratory findings and roentgenographic signs was compiled (Figs. 4 to 6). The characteristic features of each x-ray finding were identified, subdivided, and assigned a number (Figs. 2 and 3). To promote objectivity on the part of the examiners, line drawings defining certain roentgen features and terminology were prepared (Fig. 1). The roentgenograms were analyzed in a double-blind fashion with the examiners knowing only that a histologically proved ulcer was demonstrated on each. The histologic results were not available to the examiners. After the analysis had been recorded on a work sheet, the information was transferred to punched data cards to facilitate computer analysis. A computer program was employed to provide the absolute, relative, and cumulative frequency distribution of the 70 variables. The Chi-square test was used to identify those variables associated with ulcer type having a probability of 0.05 or less that the association was due to chance. Seventeen statistically valid objective variables were found.