For image denoising by wavelet transforms, it is very important how we decide the threshold. The “universal threshold” proposed by Donoho gives the optimal threshold theoretically, but the information of noise power on the image is needed. In this research, discriminant analysis, which is used for image binarization conventionally, is applied to determine optimal thresholds in wavelet domain. We selected a medical chest x-ray image of various signal-to-noise ratios for wavelet denoising. We found that our method was effective when there were many noises on the image.