We propose a cell segmentation method for analyzing images of densely clustered cells. The method combines the strengths of marker-controlled watershed transformation and a convolutional neural network (CNN). We demonstrate the method universality and high performance on three Cell Tracking Challenge (CTC) datasets of clustered cells captured by different acquisition techniques. For all tested datasets, our method reached the top performance in both cell detection and segmentation. Based on a series of experiments, we observed: (1) Predicting both watershed marker function and segmentation function significantly improves the accuracy of the segmentation. (2) Both functions can be learned independently. (3) Training data augmentation by scaling and rigid geometric transformations is superior to augmentation that involves elastic transformations. Our method is simple to use, and it generalizes well for various data with state-of-the-art performance.