[Neural network based detection of pulmonary nodules on chest radiographs]

PURPOSE: We investigated the capabilities of an artificial neural network-based Computer-Aided Diagnosis (CAD) system in improving early detection of pulmonary nodules on chest radiographs. MATERIAL AND METHODS: We used a data-set of 145 digitized chest films. Two different radiologists read the radiographs to detect the sites of possible nodules. The system uses two neural networks trained on a training-set of 100 radiographs selected from the data-set. The first network is used to focus attention on the sites of potential nodules while the second calculates the likeliness of nodule presence in ROIs. The clinical test was performed on 45 more radiographs from the training-set, but different from those in the data-set, which were positive for both benign and malignant nodules. These latter plain films showed 65 nodular lesions which differed by shape and acquisition technique. RESULTS: Sensitivity was 89% in all radiographs while specificity, evaluated by ROI, and accuracy, were 98%. CONCLUSIONS: There are potential limitations in nodule detection on plain radiographs. Some of them are operator-dependent, such as nonsystematic investigation, lesion underestimation, and poor reading, and some are technique-dependent, such as X-ray beam/tube, low voltage, patient positioning, focus-film distance and development process. CADs may contribute to improving detection of pulmonary nodules because the false-negative rate is decreased and sensitivity consequently increased. The high sensitivity and specificity rates of neural networks encourage further trials on wider data-sets to help the radiologist in the early detection of pulmonary nodules.