Wound Segmentation Network Based on Location Information Enhancement

Wound segmentation provides assistance for the diagnosis and treatment of wounds. We find that the wound image has a distinct feature, e.g., the pixel color changes gradually according to its position. Location information is essential to describe this feature. However, the current methods of wound segmentation based on deep learning have not significantly added location information into model training. In order to enhance this information, we propose a deep neural network model based on a location map and location-enhanced convolution kernel. The model effectively encodes the location information to one feature map, which is then concatenated with the inputs of the network and added to the hidden layer of the network after downsampling. Moreover, the model uses a fixed-value initialized convolution kernel to further enhance the location information in the training of the network. At the end of the model, a fixed-value depth-wise convolution layer is added to eliminate minor errors.