Deep Learning Techniques for 3D-Volumetric Segmentation of Biomedical Images

A thorough review of deep learning (DL) methods for the 3D-Volumetric segmentation of biomedical images is presented in this chapter. The performance of these deep learning methods for 3D-Volumetric segmentation of biomedical images has been assessed by the classification of these methods into tasks. We have devised two main categories, i.e., the backbone network and the task formulation. We have categorized various methods based on the backbone network into a convolutional neural network (CNN)-based and generative adversarial network (GAN)-based. These techniques are further divided into the semantic, instance, and panoptic tasks for the 3D-Volumetric segmentation of biomedical images based on the task formulation. The majority of the most prominent deep learning architectures used to segment biomedical images employ CNNs as their standard backbone network. In this field, 3D networks and architectures have been developed and put into use to fully take advantage of the contextual information in the spatial dimension of 3D biomedical images. Because of the advancements in deep generative models, various GAN-based models have been designed and implemented by the research community to address the challenging task of biomedical image segmentation. The challenges are addressed, and recommendations are provided for future studies in the domain of DL methods for 3D-Volumetric segmentation of biomedical images at the conclusion of the study. The non-local U-Net based on CNN outperforms the GAN-based FM-GAN with a Dice Similarity Coefficient (DSC) of 89% for 3D-Volumetric semantic segmentation of 6-month infant Magnetic Resonance Imaging (MRI) on the iSeg dataset. The GAN-based model MM-GAN performs better than the best CNN-based model 3D FCN with multiscale for the 3D-Volumetric semantic segmentation of the human brain, obtaining a DSC of 89.90% and 86.00% for the whole tumor (WT), respectively on the BRATS 17 dataset. For 3D-Volumetric segmentation of nuclei, RDC-Net outperforms the best GAN-based model SpCycleGAN with average precision (AP) values of 99.40% and 93.47%, respectively.