In this paper we develop a new scale adaptive scheme of wavelet thresholding for noise removal. The method uses chi-square test statistics (CTS) to discriminate between noise and signal among the wavelet coefficients. The scheme uses CTS as a ruler to measure the similarity between the statistical model and the true distribution of noise. The basic philosophy of the proposed method is similar to a recursive hypothesis testing procedure. We demonstrate this method by denoising signals corrupted with additive zero-mean Gaussian noise.