Recognition of chest radiograph orientation for picture archiving and communications systems display using neural networks

Abstract

A neural network classification scheme was developed that enables a picture archiving and communications system workstation to determine the correct orientation of posteroanterior or anteroposterior chest images. This technique permits thoracic images to be displayed conventionally when called up on the workstation, and therefore reduces the need for reorientation of the image by the observer. Feature data were extracted from 1,000 digitized chest radiographs and used to train a two-layer neural network designed to classify the image into one of the eight possible orientations for a posteroanterior chest image. Once trained, the neural network identified the correct image orientation in 888 of 1,000 images that had not previously been seen by the neural network. Of the 112 images that were incorrectly classified, 106 were mirror images of the correct orientation, whereas only 6 actually had the caudal-cranial axis aligned incorrectly. The causes for misalignment are discussed.