Boundary regression-based reep neural network for thyroid nodule segmentation in ultrasound images

Due to the noises captured in ultrasound device and image reconstruction process, the edges of thyroid nodule are usually not distinctive and it is very difficult for existing approaches to well segment them in ultrasound images. While deep neural networks like U-Net have been successfully applied in many medical image segmentation tasks, their segmentation performances on ultrasound images are still not satisfactory. To address this issue, we propose in this paper a boundary field regression branch to provide useful boundary information to help improve the segmentation performance of existing networks. Without requirement of additional labeling costs, our approach firstly generates boundary field heatmap from available segmentation masks, which are then used as a supervision to train the regression branch. As a general architecture, our branch can be integrated with all encoder-decoder like segmentation networks. A dataset consisting of 3169 images from 2004 patients is used for experiments. We integrate our branch with U-Net, Attention U-Net, U-Net++ and DeepLabv3+; consistent improvements of Dice metrics were observed. The memory and computation costs required by adding our branch are marginal as well.