Brain tumor detection with multi-scale fractal feature network and fractal residual learning

Abstract

Deep learning has enabled the creation of several approaches for segmenting brain tumors using convolutional neural networks. These methods have come about as a direct result of the advancement of the field of machine learning. The proposed pixel-level segmentation is based on fractal residual deep learning; provide an insufficient degree of sensitivity when used for tumor segmentation. This is achieved due to fractal feature extraction and multi-scale approach used for segmentation. If multi-level segmentation is used, it is possible to effectively increase the sensitivity of the segmentation process which is the additional benefit from the proposed method. In this work, the production of tumor region is based on multi-scale pixel segmentation. This approach protects the integrity of tumor information while simultaneously improving the detection accuracy by cutting down on the total number of tumor regions. When compared to the information about the brain found in tumor locations, the proposed strategy has the potential to enhance the percentage of brain tumor information. This work proposes a novel network structure known as the Mutli-scale fractal feature network (MFFN) to increase the accuracy of the network's classification as well as its sensitivity when it comes to the segmentation of brain tumors. The proposed method with overall feature results in 94.66% accuracy, 94.42% sensitivity and 92.81% specificity using 5-fold cross validation. In this paper the Cancer Imaging Archive (TCIA) dataset has been used in order to evaluate performance evaluation metrics and segmentation results to quantify the superiority of proposed brain tumor detection approach in comparison to existing methods.