U-Net for Biomedical Image Segmentation

Summary & Methodology Analysis

The paper addresses biomedical image segmentation using very few training images while achieving both precise localization and context capture. To accomplish this, the authors design a network structure featuring a contracting path and an expansive path. The contracting path uses repeated application of two 3 by 3 unpadded convolutions, each followed by a rectified linear unit (ReLU) activation function and a 2 by 2 max pooling operation with stride 2 for downsampling, doubling the number of feature channels at each step. The expansive path handles up-sampling of the feature map followed by a 2 by 2 up-convolution that halves the feature channels, concatenation with the cropped feature map from the contracting path, and two 3 by 3 convolutions each followed by a ReLU. A final 1 by 1 convolution maps each 64-component feature vector to the desired number of classes.

The training pipeline uses stochastic gradient descent in Caffe with a batch size of a single image and a high momentum of 0.99. The energy function is computed via a pixel-wise soft-max over the final feature map combined with a cross entropy loss function using a pre-computed weight map for touching cells. To handle limited training data, the authors apply excessive data augmentation using random elastic deformations generated on a coarse 3 by 3 grid with 10 pixels standard deviation. For inference on arbitrarily large images, an overlap-tile strategy mirrors input image context at borders to enable seamless segmentation without size limits.

Regarding performance, the U-Net achieves a warping error of 0.0003529 and a rand-error of 0.0382 on the EM segmentation challenge. It also achieves an average intersection over union (IOU) of 92 percent on the PhC-U373 dataset and 77.5 percent on the DIC-HeLa dataset of the ISBI cell tracking challenge 2015. The paper does not specify any computational limitations or hardware constraints.