Automatic detection of optic disc in retina image using CNN and CRF

In this paper, we propose an optic disc detection method based on convolutional neural network (CNN) and conditional random field (CRF). We pre-classify the color fundus retinal images by CNN, and construct first-order potential functions of CRF. Then the linear combination of Gaussian kernel functions is used to construct the second-order potential function of CRF model. Finally, regional restricts method is applied that analyzes the consistency of the connected region labels and corrects the labels of each pixel by calculating the posterior probability mean of the super-pixel region. The combination of CNN and CRF not only uses the pixel's intrinsic features, but also the spatial context information to make the detection more accurate. The added constraints further preserve the local information of the target and infer the entire model through a mean field approximation algorithm. This improves the accuracy of detection of optic discs in color fundus retina images. Experiments show that the CNN-CRF model performs better than the existing algorithms for the optic disc detection in pathological images. It provides an effective solution to optic disc detection problem by inhibiting its vulnerability to noise interference such as peripheral lesions and pigmentation. We compare our results to recent published results on several retina databases and show that the CNN-CRF model outperforms the current state-of-the-art methods.