Brain Tumor Segmentation and Classification Using HardNet Transformer With Gated Recurrent Unit Model

Abstract

The uncontrolled growth of brain tumor cells presents a severe health risk, making early diagnosis and accurate classification essential for improving treatment outcomes. However, the existing approaches used for segmentation fail to delineate the tumor regions accurately due to a lack of global contextual information of brain tumors, resulting in poor classification performance. To overcome this limitation, this research proposes a HardNet Transformer based segmentation and Gated Recurrent Unit (HT-GRU)-based classification model for efficient brain tumor classification. Initially, the Magnetic Resonance Images (MRI) are acquired from datasets such as Figshare, BraTS 2020, and Kaggle. Brain tumor features have complex relationships among texture, shape, and intensity features from preprocessed MRI images; therefore, a MobileNet-V2 model is utilized to capture these nonlinear relationships effectively. This approach is beneficial for brain tumor classification, where subtle variations in MRI images, such as tumor texture or boundary features, are crucial for accurate classification. Experimental results of the HT-GRU classification model achieve accuracies of 98.97 and 99.16% for the Figshare and Kaggle datasets, respectively, which are greater than those of previous approaches, such as Binomial Thresholding and Bidirectional Long Short-Term Memory (BT-Bi-LSTM).