3D HOG-Aided Convolutional Neural Network for Prostate Segmentation in Biomedical Image

Accurate and automatic segmentation of prostate magnetic resonance (MR) images is essential for the diagnosis, prediction, treatment and prognosis of prostate diseases. In this paper, we propose a three-dimensional (3D) histogram of oriented gradient (HOG)-aided convolutional neural network method to segment prostate MR images. Our model consists of three modules: 3D HOG feature extraction stage for prostate boundary information, 3D convolutional neural network (CNN) module and feature information fusion module. We extend the traditional HOG feature from two to three dimensions to provide the boundary information of prostate in MR images. To overcome the shortcomings of the traditional skip connection in the 3D-CNN module, which uses 3D-Unet as the baseline, we improved the skip connection of 3D-Unet such that each decoder layer merges the smaller and same scale feature map of the encoder with the larger scale feature map of the decoder. Finally, we perform a fusion mechanism on the two features to obtain a segmentation probability map. We evaluate our model on MR images of prostate acquired from a local hospital. The average relative volume difference (RVD) and Dice coefficient (DSC) of the prostate volume area are 1.90% and 90.83%, respectively. Quantitative experiments show that, compared with other methods, the results of our method are highly consistent with expert manual segmentation.