Tahereh Hassan Zadeh Koohi1, Len Hamey2, Kevin Ho-Shon3
1 Department of Computing, Macquarie University, Balaclava Road, North Ryde, Sydney, NSW, 2109, Australia, tahereh.hassan-zadeh-koohi@hdr.mq.edu.au
2 Department of Computing, Macquarie University, Balaclava Road, North Ryde, Sydney, NSW, 2109, Australia, len.hamey@mq.edu.au
3 Department of Computing, Macquarie University, Balaclava Road, North Ryde, Sydney, NSW, 2109, Australia, kevin.ho-shon@mq.edu.au
Digital medical image segmentation is the process of partitioning an image into several discrete and homogeneous regions. Segmentation is needed to find the boundary of the prostate either automatically or semi-automatically. One of the most accurate and non-invasive prostate imaging methods is Magnetic Resonance Imaging (MRI) which is usually employed for the prostate image segmentation and/or possible prostate anomalies detection.
In this research, to improve the Fully Convolutional Neural Network (FCNN) performance for prostate MRI segmentation, we analyse various structures of shortcut connections as well as the size of a deep network. We suggest eight different deep 2D network structures for automatic MRI prostate segmentation based on FCNN. Our evaluations on the PROMISE12 dataset with ten-fold cross-validation indicate improved and competitive results. We analyse the results in detail, considering MRI slices, MRI volumes, test folds, and also the impact on prostate segmentation of using an EndoRectal Coil to capture the prostate MRI. Our best 2D network outperforms the state-of-the-art 3D FCNN-based methods for prostate MRI segmentation, without any further post-processing module nor pretraining on publicly available data.