Microcalcification detection for mass screening programmes

Discusses some of the issues involved in the selection of appropriate metrics of the evaluation of microcalcification detection algorithms. To measure the sensitivity the authors use a database consisting of films containing clusters associated with biopsy-proven malignancies. In order to measure the algorithm´s specificity the authors use a database of non-recalled film pairs. The adoption of these databases overcomes some of the problems associated with the grey area of biopsy-proven benign recalls. The authors quote their results on a per woman basis because this reduces the influence of the variability of radiologist´s annotations and also, they believe, gives a better insight into the algorithm´s likely effect within a prompting environment. Results detailing the causes of the false prompts within the database used to measure the specificity are presented. These correspond to a fairly high sensitivity measure. The authors believe that these results will be typical of what can be expected when an algorithm is run on large batches of unbiased films, the vast majority of which a radiologist would not recall. From these results the authors conclude that the classification of some clusters as bring clearly benign is an important step towards a successful prompting system. Of all these clearly benign calcification structures vascular calcifications appear to be by far the most frequent and thus an algorithm which correctly classifies them as such is highly desirable. Finally the authors discuss the difference between an algorithm´s ROC graph and that of a radiologist prompted by that algorithm. In theory there may be little correspondence between the two. The authors discuss some of the practical difficulties involved in producing the ROC graph for a prompted radiologist and suggest the use of library films of missed cancers may be helpful in reducing the range of parameters sets used to product this ROC graph