Neural Networks in Radiologic Diagnosis; II. Interpretation of Neonatal Chest Radiographs

Abstract

A neural network (NN) system was trained to choose one or more diagnoses from a list of 12 possible diagnoses, based on 21 radiographic observations made on each of a series of neonatal chest radiographs. Initially, an experienced pediatric radiologist provided both the radiographic observations and ranked differential diagnoses for each of 77 neonatal chest radiographs in the preliminary phase used to train the NN. Subsequently, two pediatric radiologists (one of whom provided the initial training-phase data) independently read a series of 103 neonatal chest radiographs (different from the training set) and compiled a list of radiographic findings and differential diagnoses for each radiograph. The trained NN was then asked to provide a list of differential diagnoses for each case from the radiologists' lists of findings. Agreement between the network and each radiologist independently was greater than between the two radiologists. Both the positive and negative agreement between the network and either radiologist was greater than the inter-radiologist agreements for most of the diagnostic endpoints.