Abstract
Reliable and automated segmentation of the prostate in magnetic resonance (MR) images is critical for the diagnosis and treatment of prostate cancer. The unclear boundary of the prostate increases the difficulty of correctly differentiating the prostate region from the surrounding regions. This paper proposes a novel approach with Residual and Gated Network (ResGNet) for prostate segmentation on MR images. The ResGNet block fuses the residual and gated connections in a multi-directional and multi-path learning strategy for enhancing information propagation and preserving important boundary information. The proposed method is conducted on several public databases, and an average DSC of 94.4%, 95-HD of 3.280 mm, and ABD of 0.919 mm are obtained in the 5-fold cross-validation experiment on the PROMISE12 dataset. Experimental results demonstrate that the ResGNet outperforms the state-of-the-art methods and is effective and robust regardless of the variability of MR scanners and acquisition protocols or the size and shape of the prostate gland for the segmentation of healthy or diseased prostates.
Original language | English (US) |
---|---|
Article number | 105508 |
Journal | Biomedical Signal Processing and Control |
Volume | 87 |
DOIs | |
State | Published - Jan 2024 |
Externally published | Yes |
Keywords
- Gated convolutional neural network
- Magnetic resonance imaging
- Prostate cancer
- Prostate segmentation
- Residual convolutional neural network
ASJC Scopus subject areas
- Signal Processing
- Biomedical Engineering
- Health Informatics