A concise review of U-Net-based deep learning models for brain tumor segmentation

Abstract

      A brain tumor requires highly discriminative segmentation and classification of the tumor in order to provide effective treatment. Three categories of brain tumor segmentation (BTS) exist fully automated, semi-automatic and manual. Both tumor segmentation in therapy, treatment planning, and diagnostic assessment have frequently relied on the deep learning (DL) approach to automate the process. It is founded on the U-Net paradigm that has recently demonstrated a superb performance in multimodal BTS.  The study retrieves a literature review with the application of U-Net models to the BTS.  It is a general method of training a new U-Net model on brain tumor segmentation. The procedures of this DL method have been explained to allow the derivation of the desired model. These involves collection of the dataset, pre-processing, selection/ or development of structure of model, transfer learning (optional) and image enhancement (optional). The performance and structure of the model are the two most essential measures that are used to assess the literature.  Based on the findings of the review, there is a direct relationship between model accuracy and the architectural complexity of that model; therefore, it will be a future challenge to produce increased accuracy using a less complex design.  There are also futures trends, alternatives, and challenges.