PROJECT TITLE :

3D APA-Net 3D Adversarial Pyramid Anisotropic Convolutional Network for Prostate Segmentation in MR Images

ABSTRACT:

Diagnostic and treatment of prostate illnesses, particularly cancer, rely heavily on accurate and reliable prostate gland segmentation utilising MR imaging. It's still possible to increase the performance of automatic segmentation based on deep learning despite a wide range of picture appearances, imaging interference, and anisotropic spatial resolution. For prostate segmentation in MR images, we present the 3D adversarial pyramid anisotropic convolution neural network (3D APA-Net). An image generator (i.e., 3D PA-Net) and a discriminator (i.e., a six-layer convolutional neural network) are used in this model to perform image segmentation and distinguish between a segmentation result and its ground truth. Encoding and decoding are done in 3D using the ResNet encoder and an anisotropic convolutional decoder using multi-level pyramid convolutional skip connections in the 3D PA-Net. Convolutional blocks with anisotropic resolution, pyramid convolutional blocks, and adversarial training regularise the 3D PA-Net and hence enable it to produce spatially consistent outputs. It was tested on two public databases and a hybrid of the two to see how well the planned 3D APA-Net performed. As a clinical workflow, the suggested model beats the compared approaches on three databases, according to our findings.