[Preliminary clinical evaluation of computer-aided diagnosis in digital chest radiography]

We are developing computer-aided diagnosis (CAD) schemes for chest radiography with which to alert radiologists to possible lesions, and thus potentially improve diagnostic accuracy. Although preliminary results of individual CAD schemes have been encouraging, CAD is still at an early stage of its development. Therefore, we applied our CAD programs in a clinical environment and evaluated its potential usefulness and limitations. CAD programs including automated detection of cardiomegaly, analysis of interstitial infiltrates, and detection of lung nodules were applied to 310 consecutive chest radiographs. CAD results for cardiomegaly and interstitial infiltrates were evaluated subjectively by radiologists and physicists for clinical accuracy and technical problems, respectively, depending on the correspondence of the CAD output to the nature and location of important findings influencing the radiologic diagnosis. CAD results for lung nodules were evaluated objectively by taking into account the number of false positives and true positives. For detection of cardiomegaly, 71% of evaluated cases were judged to be acceptably accurate in terms of clinical accuracy. For analysis of focal and diffuse interstitial infiltrates, 87% and 61% were acceptably accurate, respectively. For detection of lung nodules, 61% were acceptably accurate. More than 70% of CAD output was judged to be possibly accurate in terms of technical aspects. Technical problems were related to failure to detect subtle abnormalities and the occurrence of false positives caused by normal anatomical structures. We believe that CAD has the potential to be a valuable aid for radiologists in clinical practice, if certain technical problems can be overcome and if optimal operating points can be defined for clinical use.