Segmentation of brain tumor subregions with depthwise separable dense U-NET (DSDU-NET)

Brain tumors are one of the most dangerous medical conditions. The dispersed extremities and nonuniform structure of the tumors are the basis of why techniques of traditional segmentation have grown to be inefficient. Magnetic resonance imaging (MRI) is one of the most widely used scanning procedures for tumors. However, to ameliorate the survival rate, the detection of tumors alone does not suffice, and there are other effective procedures. One of the most pivotal procedures for diagnosing the condition is the process of the brain tumor segmentation is a laborious and cumbersome process. As a result, a deep learning (DL) based solution is used to extract tumor subregions like enhancing tumor (ET), tumor core (TC), and whole tumor (WT). The proposed model involves novel DSDU-Net: Depth-wise separable dense U-NET (DSDU-NET) that precise outputs are acquired by retaining the low-level features. In the preprocessing stage, a methodology called multi-scale patch extraction is used to segregate tumor regions, and a grouping of Gaussian filter, unsharp masking, and histogram equalization is carried out on the data set to get significantly better performance. The proposed DSDU-NET network performed better in terms of performance, specificity, sensitivity, dice similarity index, and Hausdorff distance for segmented image sub-regions when validated on BraTS 2018 and 2019 data sources when particularly in comparison to other models.