The diagnosis, assessment, and prognosis of prostate disorders (including prostate cancer) depend heavily on the automatic segmentation of prostate magnetic resonance (MR) images. Convolutional neural networks have recently replaced traditional prostate segmentation techniques. The segmentation performance should be enhanced further due to the intricacy of the tissue structure in MR images and the limits of the current approaches to spatial context modeling.
For a study, researchers suggested a brand-new 3D pyramid pool Unet (3D PPU-net), which makes use of the pyramid pooling structure inherent in the skip connection and the deep supervision in the 3D Unet’s up-sampling. The typical 3D Unet network’s parallel skip connection sends low-resolution data to the feature map repeatedly, blurring the features of the images. They combined each decoder layer with the feature map of the same scale as the encoder and the smaller-scale feature map of the pyramid pooling encoder in order to address the drawbacks of the standard 3D Unet. The low-level specifics and high-level semantics at 2 separate layers of feature maps were combined in the skip link. In addition, deep supervision learns hierarchical representations from thorough aggregated feature maps, which can increase task accuracy. Pyramid pooling also does multi-faceted feature extraction on each picture behind the convolutional layer.
Research on 78 patients’ 3D prostate MR images showed that the findings were significantly linked with manual segmentation performed by a skilled expert. The prostate volume region had an average relative volume difference (RVD) and dice similarity coefficient (DSC) of 2.32% and 91.03%, respectively.
Quantitative tests showed that the method’s outcomes were significantly more consistent with the expert manual segmentation when compared to those of other approaches.
Reference: aapm.onlinelibrary.wiley.com/doi/10.1002/mp.15895
