抄録
We propose an automatic light spot detection method in intracellular images using contextual information. Light spot detection in intracellular images is important for classification of patients condition. However, light spots are detected manually. Thus, we propose a light spot detection method in intracellular images by computer. In general, supervised learning is used to develop a light spot detector with high accuracy. However, there are the cases in which large number of supervised data are not obtained. In fact, in our research, we have only 4 images with supervised signals. Therefore, we use background subtraction and robust statistics to detect light spots because they do not require supervised signals. However, only unsupervised learning does not give high accuracy. In particular, we can not discriminate light spot and noise well. Therefore, we use contextual information which is obtained from neighboring region of a light spot to reduce noise. In experiments, our method is applied to Wnt-3a images and achieved 80.71% which is higher accuracy than a supervised detector with small number of training images.
