Collimation field detection is an important pre-processing step for automatic image analysis of radiographs. However, most approaches are restricted to a small set of form archetypes or presuppose the presence of a shutter. Hence, existing methods are not applicable to large collections of radiographs from various modalities, such as obtained in the field of content-based image retrieval in medical applications. Based on analytical evaluation, the approach of WIEMKER et al. (Procs SPIE 2000; 3979:1555-1565) was selected, modified in order to reduce false positive detection, and evaluated on a large set of 4,000 radiographs (763 containing shutter edges) taken from daily routine including any kind of projective X-ray examinations. Eight subsets (each of 500 images) were compiled randomly. A set of 500 images was used to optimize the parameters and evaluated using the remaining 3,500 images. This procedure was repeated for all eight combinations. Using the initial approach, the specificity is 96.4% with a poor sensitivity of 44.1% resulting in an overall precision of 86.7%. All figures increase up to 98.5%, 55.6%, and 89.5%, respectively, if the algorithm also minimizes the variation of radiation density values outside the detected shutter area. In terms of sensitivity and precision, the results of optimization vs. evaluation for the same combination and of evaluation vs. evaluation for different combinations differed up to 13 and 9 percentage points, respectively. This indicates that still an insufficient number of images is used to allow complete generalization of the results.