Use of an artificial neural network to determine the diagnostic value of specific clinical and radiologic parameters in the diagnosis of interstitial lung disease on chest radiographs

Rationale and Objectives. The authors investigated the diagnostic value of each of multiple clinical parameters and radiologic findings in differentiating between various interstitial lung diseases by using an artificial neural network (ANN). Materials and Methods. The ANN was designed to differentiate between 11 interstitial lung diseases. The authors employed 10 clinical parameters and 16 radiologic findings that were divided into three groups (location, general appearance, specific findings). The performance of the ANN was evaluated with receiver operating characteristic analysis with a modified round-robin (leave-one-out) method and 370 cases (150 actual cases, 110 published cases, and 110 hypothetical cases). The A, values of ANNs were evaluated with various combinations of 10 clinical parameters and 16 radiologic findings. Results. The A, value obtained with the complete set of clinical parameters and radiologic findings was 0.947. The A, value obtained with the 10 clinical parameters alone was 0.900, which was greater than 0.843 obtained with the 16 radiologic findings alone. There were statistically significant differences among A., values for some diseases when certain clinical parameters were removed from the input. Omission of specific findings among the three groups of radiologic findings decreased the A, value significantly. Conclusion. These results appear to confirm that clinical parameters can be equally as or more important than radiologic findings in the diagnosis of interstitial lung diseases. Among radiologic findings, certain specific findings can be more important than the location or general appearance of abnormal findings.