Learning-based method for p53 immunohistochemically stained cell image segmentation

In this study, a learning-based color image conversion method is proposed for cell image segmentation. Firstly, we demonstrate that minimum distance-based pixel classification, such as clustering, for color image segmentation in the color space is equivalent to thresholding grayscale images. Motivated by this result, we develop the so called C-G-T procedure for color image segmentation, where color image (C) is first converted into grayscale (G) and thresholding (T) is then performed on the gray image to segment objects out of background. The transform for image conversion is learned from the global pixel distribution in the color space, while the threshold is learned from local pixel distribution of the gray image. The combination of global and local learning makes the C-G-T procedure adaptive and computational efficient. Extensive experiments are performed to verify the effectiveness of our method.