The image coding algorithm “Set Partitioning in Hierarchical Trees (SPIHT)” introduced by Said and Pearlman achieved an excellent rate-distortion performance by an efficient ordering of wavelet coefficients into subsets and bit plane quantization of significant coefficients. We observe that there is high correlation among the significant coefficients in each SPIHT pass. Hence, in this paper we propose trained scalar-vector quantization (depending on a boundary threshold) of significant coefficients to exploit correlation. In each pass, the decoder reconstructs coefficients with scalar or vector quantized values rather than with bit plane quantized values. Our coding method outperforms the scalar SPIHT coding in the high bit-rate region for standard test images.