Most of lossy image compression methods like SPIHT or JPEG2000 are based on determining the intervals in which each coefficient can be reached. Once these intervals are known, the decoder selects the value that minimizes the mean square error of the possible values. When a coefficient is not considered significant, because its sign is not known, the selected reconstruction value is zero. This ignorance of the sign for non significant coefficients makes the interval of the possible values to be twice as much the rest of the Wavelet coefficients. In this paper an alternative for the selection of the reconstruction value for the non significant coefficients is analyzed. It is based on predicting their actual values using the neighbors' values. With this new improvement it has been possible to reach average reductions of about 5 percent, being higher than reductions reached by JPEG2000 over SPIHT.