Brain tumor classification via UNet architecture of CNN technique

Abstract

The key step of early diagnosis in the brain tumor classification highly depends on the accurate estimation through the CT images or the other datasets. The precise results can effectively enhance the survival rate of patients. However, it becomes a difficult task to achieve a robust classification model with the help of traditional deep learning approaches due to differences in texture, shape and size of tumor masses with visual similarities. Therefore, this paper proposed a new algorithm of UNET that address such an important issue with the three core ideas as: 1) the construction of U-shaped classification framework for brain tumor recognition, 2) integration of adaptive histogram equalization approach with the suggested truncated normalization method for the enhancement of image contrast and 3) transfer learning approach to sort out the issue of insufficient training data. The proposed scheme shows superior performance by evaluating the public datasets of brain tumor. UNET architecture becomes a trending topic among the researchers and broadly utilized as a tool for medical image segmentation. Although, the conventional level set approaches still remained an issue as experts fail to completely understand the tumor regions due to their low-level features with complex characteristics. Thus, it is important to consider the contextual features that can deliver the complementary discriminative data to low-level characteristics for brain tumor segmentation. It starts with the designing of an UNET algorithm for encoder-decoder framework that acquire the knowledge of semantic information as high-level contextual features. Afterwards, the outcomes depicts the novel contextual energy terms with the potential to integrate the high-level contextual information in the level set framework. Such semantic information with learned contextual features can deliver the accurate distinctive information that can directly connect the category labels rather with original intensity. Thus, it can enhance the segmentation performance as it becomes robust to serious intensity inhomogeneity.