To reduce inter- and intra-reader variability in diagnosing chest radiographs, a neural network-based computer-aided diagnostic system was developed and tested. The results of an experiment with 124 digitized chest radiographs, demonstrated high degrees of sensitivity and specificity in classifying chest radiographs. The use of a computer-assisted chest radiograph reader eliminated the inconsistencies in the human readers. The Computer-assisted Chest Radiograph Reader System (CARRS) applies recognized principles in the psychophysics of human vision, incorporates neural network-based image analysis and integrates these with a graphical user interface. Advances in digital image processing, and classification techniques have made CARRS feasible for meeting screening, research and development, and clinical requirements. Through the adoption of the International Labor Organization (ILO) classification procedures, it had been hoped that reader variation in the classification of parenchymal abnormalities could be minimized. The ILO classification of the pneumoconioses is based on a structured procedure for detecting and characterizing patterns on chest radiographs. Numerous studies have shown, however, that inter- and intra-observer variability of radiograph readings by trained medical personnel has persisted. The methodology was implemented through the following tasks: 1) From a data base of several thousand patients, a set of 205 chest radiographs were manually graded by two pulmonologists; 124 of the films were then digitized at 12-bit high spatial resolution. 2) Textural features were calculated using high order statistical techniques. The features were classified by the pulmonologists to "train a neural network to extract classification rules chest radiographs based on the ILO methodology. 3) The neural network classification from the graded system was tested using 65 chest radiographs. For 5-10 areas selected by the pulmonologist on the chest radiograph, a feature vector composed of image characteristics such as density distribution, entropy, fractal dimension, opacity counts, shape, etc. was calculated. This feature vector characterized numerically the areas used by the pulmonologist to grade the radiograph. The neural network trained on the same regions used by the pulmonologist, and through a quantitative feature vector, "learned" the characteristics of each ILO classification.
Kestrel Corporation, 6624 Gulton Court NE, Albuquerque, NM 87109