The automatic detection of flaws through non-destructive testing uses pattern recognition methodology with binary classification. In this problem a decision is made about whether or not an initially segmented hypothetical flaw in an image is in fact a flaw. Neural classifiers are one among a number of different classifiers used in the recognition of patterns. Unfortunately, in real automatic flaw detection problems there are a reduced number of flaws in comparison with the large number of non-flaws. This seriously limits the application of classification techniques such as artificial neural networks due to the imbalance between classes. This work presents a new methodology for efficient training with imbalances in classes. The premise of the present work is that if there are sufficient cases of the smaller class, then it is possible to reduce the size of the larger class by using the correlation between cases of this latter class, with a minimum information loss. It is then possible to create a training set for a neural model that allows good classification. To test this hypothesis a problem of great interest to the automotive industry is used, which is the radioscopic inspection of cast aluminium pieces. The experiments resulted in perfect classification of 22936 hypothetical flaws, of which only 60 were real flaws and the rest were false alarms.