A two-stage segmentation of sublingual veins based on compact fully convolutional networks for Traditional Chinese Medicine images.

Abstract

As one of the key methods of Traditional Chinese Medicine inspection, tongue diagnosis manifests the advantages of simplicity and directness. Sublingual veins can provide essential information about human health. In order to automate tongue diagnosis, sublingual veins segmentation has become one important issue in the field of Chinese medicine medical image processing. At present, the primary methods for sublingual veins segmentation are traditional feature engineering methods and the feature representation methods represented by deep learning. The former, which mainly based on colour space, belongs to unsupervised classification method. The latter, which includes U-Net and other deep neural network models, belongs to supervised classification method. Current feature engineering methods can only capture low dimensional information, which makes it difficult to extract efficient features for sublingual veins. On the other hand, current deep learning methods use down-sampling structures, which manifest weak robustness and low accuracy. So, it is difficult for current segmentation approaches to recognize tiny branches of sublingual veins. To overcome the above limits, this paper proposes a novel two-stage semantic segmentation method for sublingual veins. In the first stage, a fully convolutional network without down-sampling is used to realize the accurate segmentation of the tongue that includes the sublingual veins to be segmented in the next stage. During the tongue segmentation, the proposed networks can effectively reduce the loss of medical images spatial feature information. At the same time, in order to expand the receptive field, the dilated convolution has been introduced to the proposed networks, which can capture multi-scale information of segmentation images. In the second stage, another fully convolutional network has been used to segment the sublingual veins on the base of the results from the first stage. In this model, proper dilated convolutional rates have been selected to avoid gridding issue. In order to keep the quality of the images to be segmented, several particular data pre-processing and post-processing have been used, which includes specular highlight removal, data augmentation, erosion and dilation. Finally, in order to evaluate the performance of the proposed model, segmentation results have been compared with the state-of-the-art methods on the base of the dataset from Shanghai University of Traditional Chinese Medicine. The effectiveness of sublingual veins segmentation has been proved.