We propose a method for colorization of medical grayscale images using color learning. The colors are learned from a color image and predicted for a grayscale image. Earlier we introduced an efficient algorithm for image colorization which uses a dichromatic reflection model. The colorization algorithm is further developed in this study. First, we improve the algorithm performance by extending its capability to work with the grayscale images the contrast of which is lower than the contrast of the color images. Then, we propose a reliable technique to prevent negative contrast during colorization. In addition, we develop a simple approach for grayscale image colorization by a given RGB value. We give two medical applications of our algorithm: realistic color labeling of skin wounds and colorization of a dental cast models. In the former case we use grayscale images and labeling obtained after support vector classification as input data and for the latter application we use photometric stereo images.