Radiological and routine autopsy findings from 241 South African gold miners were compared, using pathology as the "gold standard." Previous annual screening miniradiographs were read independently by two readers, using the International Labor Office (ILO) grading system, without reference to personal identifiers. Individual and consensus silicosis grades were recorded for each subject. When pathological and radiological silicosis were defined as any abnormal grade, the sensitivity and specificity of the radiological diagnosis was 67.5% and 80%, with positive and negative predictive values of 63% and 83%. Most undetected autopsy silicosis was early-grade. Using higher pathological and radiological grades to define silicosis (5 nodules or more and ILO grades 1/1 and above), sensitivity and specificity increased to 71% and 96%, and positive and negative predictive values increased to 76% and 95%, respectively. Use of a consensus grade made little difference to results from individual readers. For each radiological definition, sensitivity was considerably higher than, but specificity was similar to, that found in a previous study of South African gold miners which used the same pathology source and standard sized films. The difference between these two study findings, and unexpected demonstration of higher sensitivity from miniradiographs, suggests that further research is required into factors affecting radiological interpretation before silicosis grading can be considered to be adequately standardized.