ResUNet+: A New Convolutional and Attention Block-Based Approach for Brain Tumor Segmentation

Abstract

The number of brain tumor cases has increased in recent years. Therefore, accurate diagnosis and treatment of brain tumors are extremely important. Accurate detection of tumor regions is difficult, even for experts, because brain tumor images are low-contrast, noisy and contain normal tissue-like structures. Therefore, in this study, a new convolution-based hybrid model was proposed to perform segmentation with high accuracy. In the proposed model, instead of applying convolution to the whole image, convolution was applied to the ROI regions detected in different modalities. With this approach, it was determined that the processing cost is reduced, and the performance is increased. The proposed model was tested on BraTS 2020, BraTS 2019, and BraTS 2018 datasets. The proposed method in the study was also compared with SOTA methods using the same dataset. As a result of the comparison, dice scores of 92.80%, 93.10%, and 91.90% were respectively obtained for whole tumors, enhance tumors and tumor nuclei in the images in the BraTS 2020 dataset. With these results, the proposed model can compete with many models in the literature using the same datasets. The proposed model is a new method that can be preferred in different segmentation applications due to its performance success and especially the advantage of the pre-processing structure.