This study aimed to determine the reliability and accuracy of the ICDAS and radiographs in detecting and estimating the depth of proximal lesions on extracted teeth. The lesions were visible to the naked eye. Three trained examiners scored a total of 132 sound/carious proximal surfaces from 106 primary teeth and 160 sound/carious proximal surfaces from 140 permanent teeth. The selected surfaces were first scored visually, using the 7 classes in the ICDAS. They were then assessed on radiographs using a 5-point classification system. Reexaminations were conducted with both scoring systems. Teeth were then sectioned and the selected surfaces histologically classified using a stereomicroscope (×5). Intrareproducibility values (weighted kappa statistics) for the ICDAS for both primary and permanent teeth were >0.9, and for the radiographs between 0.6 and 0.8. Interreproducibility values for the ICDAS were >0.85, for the radiographs >0.6. For both primary and permanent teeth, the accuracy of each examiner (Spearman's correlation coefficient) for the ICDAS was =0.85, and for the radiographs =0.45. Corresponding data were achieved when using pooled data from the 3 examiners for both the ICDAS and the radiographs. The associations between the 2 detection methods were measured to be moderate. In particular, the ICDAS was accurate in predicting lesion depth (histologically) confined to the enamel/outer third of the dentine versus deeper lesions. This study shows that when proximal lesions are open for inspection, the ICDAS is a more reliable and accurate method than the radiograph for detecting and estimating the depth of the lesion in both primary and permanent teeth.