Application of an artificial neural network in radiographic diagnosis.

The description of 44 cases of bone tumors was used by an artificial neural network to rank the likelihood of 55 possible pathologic diagnoses. The performance of the artificial neural network was compared with the performance of experienced (3 or more years of radiology training) residents and inexperienced (less than 1 year of radiology training) residents. The artificial neural network was trained using descriptions of 110 radiographs of bone tumors with known diagnoses. The descriptions of a separate set of 44 cases were used to test the neural network. The neural network ranked 55 possible pathologic diagnoses on a scale from 1 to 55. Experienced and inexperienced residents also ranked the possible diagnoses in the same 44 cases. Inexperienced residents had a significantly lower mean proportion of diagnoses ranked first or second than did the neural network. Experienced residents had a significantly higher proportion of correct diagnoses ranked first than did the network. Otherwise, a significant difference between the performance of the network and experienced or inexperienced residents was not identified. These results demonstrate that artificial neural networks can be trained to classify bone tumors. Whether neural network performance in classification of bone tumors can be made accurate enough to assist radiologists in clinical practice remains an open question. These preliminary results indicate that further investigation of this technology for interpretation assistance is warranted.