Fast medical image segmentation based on patch sharing

The lack of labeled medical data is a severe challenge of applying CNNs in medical image segmentation. The common method to solve this problem is employing patches extracted from every pixel of the entire image as train samples. But classifying every pixel in the image is time-consuming, which is not appropriate in practical medical application. This paper proposed a fast segmentation algorithm based on trained network model to reduce test time. Transforming the fully-connected layer of trained network into convolutional layer is used as the test network and the entire image is the input of test network. However, due to the convolutional and pooling operation of CNN, some pixel classification results are missed. To obtain corresponding segmentation, different size of original image are cropped as the input of test network, and interpolation is taken to supply the final image segmentation, according to the offset rule of the input images. The numerical simulation experiments indicated that the proposed algorithm show prominent performance in segmentation time and remain unchanged in the final segmentation result compared with initial train network architecture.