BACKGROUND: Computed tomography (CT) is the benchmark for diagnosis emphysema, but is costly and imparts a substantial radiation burden to the patient. OBJECTIVE: To develop a computer-aided procedure that allows recognition of emphysema on digital chest radiography by using simple descriptors of the lung shape. The procedure was tested against CT. METHODS: Patients (N=225), who had undergone postero-anterior and lateral digital chest radiographs and CT for diagnostic purposes, were studied and divided in a derivation (N=118) and in a validation sample (N=107). CT images were scored for emphysema using the picture-grading method. Simple descriptors that measure the bending characteristics of the lung profile on chest radiography were automatically extracted from the derivation sample, and applied to train a neural network to assign a probability of emphysema between 0 and 1. The diagnostic performance of the procedure was described by the area under the receiver operating characteristic curve (AUC). RESULTS: AUC was 0.985 (95% confidence interval, 0.965-0.998) in the derivation sample, and 0.975 (95% confidence interval, 0.936-0.998) in the validation sample. At a probability cutpoint of 0.55, the procedure yielded 92% sensitivity and 96% specificity in the derivation sample; 90% sensitivity and 97% specificity in the validation sample. False negatives on chest radiography had trace or mild emphysema on CT. CONCLUSIONS: The computer-aided procedure is simple and inexpensive, and permits quick recognition of emphysema on digital chest radiographs.
BACKGROUND: Computed tomography (CT) is the benchmark for diagnosis emphysema, but is costly and imparts a substantial radiation burden to the patient. OBJECTIVE: To develop a computer-aided procedure that allows recognition of emphysema on digital chest radiography by using simple descriptors of the lung shape. The procedure was tested against CT. METHODS:Patients (N=225), who had undergone postero-anterior and lateral digital chest radiographs and CT for diagnostic purposes, were studied and divided in a derivation (N=118) and in a validation sample (N=107). CT images were scored for emphysema using the picture-grading method. Simple descriptors that measure the bending characteristics of the lung profile on chest radiography were automatically extracted from the derivation sample, and applied to train a neural network to assign a probability of emphysema between 0 and 1. The diagnostic performance of the procedure was described by the area under the receiver operating characteristic curve (AUC). RESULTS: AUC was 0.985 (95% confidence interval, 0.965-0.998) in the derivation sample, and 0.975 (95% confidence interval, 0.936-0.998) in the validation sample. At a probability cutpoint of 0.55, the procedure yielded 92% sensitivity and 96% specificity in the derivation sample; 90% sensitivity and 97% specificity in the validation sample. False negatives on chest radiography had trace or mild emphysema on CT. CONCLUSIONS: The computer-aided procedure is simple and inexpensive, and permits quick recognition of emphysema on digital chest radiographs.