This report describes a mathematical approach for classifying two-dimensional (2D)-protein maps without spot detection or pattern matching. Analysis of electrophoretograms was performed using wave packet decompositions of the signals. The scanned images were automatically decomposed into a set of sub-images organized in a tree structure. Each sub-image contained relevant information such as its energy, or entropy. Moreover the node position itself of the sub-image reflected a frequency localization. A distance was then defined using the tree repartition of these quantities. Finally a statistical clustering on the tree structures was performed, terminating with a classification of the images according to their repartition frequencies. The algorithm has been applied to classify immunoglobulin (Ig) light chain patterns and proved useful to automatically detect monoclonal, oligoclonal or polyclonal Igs.