Highly Efficient Forward and Backward Propagation of Convolutional Neural Networks for Pixelwise Classification