3D Fully Convolutional Network Incorporating Savitzky-Golay Filtering for Prostate Segmentation

In this paper, we proposed a 3D fully convolutional network (FCN) incorporating Savitzky-Golay (SG) filtering for prostate segmentation using magnetic resonance images (MRIs). Deep learning methods have achieved promising results in the field of segmentation, especially in semantic segmentation. However, it is not fully applicable to 3D medical images. To better extract the spatial information encoded in the 3D volumetric data, we designed a 3D FCN with long skip connection and the Parametric Rectified Linear Unit (PReLU) being the activation function. To further polish the deep learning based segmentation results, we employed SG filtering as a post-processing step. The SG filter was applied for smoothing and denoising, wherein second-order partial derivatives were taken to extract the edge information and achieve hole filling. In comparison with the 3D FCN without SG filtering, the post-processed results were more smooth, accurate and robust. The proposed method performed superiorly for prostate segmentation over several other state-of-the-art methods.