Figure 1: (A) Original images. (B) Seed detection-based cell segmentation. The Euclidean distance transform map is first calculated from the binary image, and single-pass voting is employed to localize seed candidates. Finally mean shift is applied to final seed detection (green crosses). With the detected seeds, a repulsive balloon snake model is used to segment individual cells (green contours). (C) For each segmented cell, three types of cellular features are calculated. To describe the whole image patch, we mapped each feature into 6-dimensional vectors for geometry and intensity features. We computed 4 statistical moments for textures before the mapping is performed. (D) Image based survival model to predict survival. 82 training data were used to build final model and select significant features as well as determine cut-off point for stratifying risk group. 40 testing set were applied to evaluate model performance by proposed criteria.