Background: The high incidence rate of prostate disease poses a requirement in accurate early detection. However, as one of the main imaging methods used for prostate cancer detection, Magnetic Resonance Imaging (MRI) has appearance variation and imbalanced problems, so automated prostate segmentation is still challenging.
Objective: Aiming to segment the prostate accurately from MRI, we focused on designing a unique network with benign loss functions.
Methods: We proposed a novel Densely Dilated Spatial Pooling Convolutional Network (DDSP ConNet) in encoder-decoder structure with a unique DDSP block. By densely combining dilated convolution and global pooling layers, the DDSP block supplies coarse segmentation results and preserves hierarchical contextual information. Meanwhile, we adopted DSC and Jaccard loss to train our DDSP ConNet. And we theoretically proved that, they have benign properties, including symmetry, continuity and differentiability about the parameters of network.
Results: To corroborate the effectiveness of our DDSP ConNet with DSC and Jaccard loss, extensive experiments have been conducted on the MICCAI
PROMISE12 challenge dataset. In the test dataset, our DDSP ConNet achieved a score of 85.78.
Conclusion: In our experiments, the DDSP network with DSC and Jaccard loss outperformed most of other competitors in the PROMISE12 dataset. Therefore, it has better ability in extracting hierarchical features and addressing imbalanced medical image problem.