This paper proposes an adaptive intra prediction method for DCT-based image coding. In this method, predicted values in each block are generated in spatial domain like the conventional intra prediction methods. On the other hand, prediction residuals to be encoded are separately calculated in DCT domain, i.e. differences between the original and predicted values are calculated after performing DCT. Such a prediction framework allows us to change the coding process from block-wise order to coefficient-wise one. When the coefficient-wise order is adopted, a block to be predicted is almost always surrounded by partially reconstructed image signals, and therefore, efficient interpolative prediction can be performed. Simulation results indicate that the proposed method is beneficial for removing inter-block correlations of high-frequency components.