Extension Network of Radiomics-Based Deeply Supervised U-Net (ERDU) for Prostate Image Segmentation

Automatic prostate segmentation from MRI images is important in disease diagnosis and treatment. The main challenges are the complex boundaries, the spatial and morphological heterogeneity, and the variety of prostate shapes. This paper proposes a deep CNN network based on 2D Res-Unet with equalization and noise reduction for preprocessing using a median filter. Additionally, a residual connection and batch normalization are used in the UNet-based network to improve gradient flow and avoid overfitting the network. The 2D Res-Unet method showed promising results on the PROSTATEx prostate MRI dataset. It achieves a dice similarity coefficient of 82.7% with a small number of parameters while outperforming the standard benchmark algorithms. Our results show that the EDRU network achieves more accurate results than the state-of-the-art U-net network for prostate gland segmentation.