HRNet: Hybrid Residual Network and Spatial Attention-based Bilateral Network with Weighted FeatureSelection for Brain Tumor Segmentation andClassification

Introduction: The occurrence of abnormal cells in the brain causes brain tumours. Magnetic Resonance Imaging (MRI) is usually employed to detect brain tumors in humans. The unusual tissue growth in the brain is determined through the information from the MRI. Machine learning methods are utilized in various research works for the localization of brain tumours. Tumor detection is rapidly performed when the approaches are applied to the MRI. However, segmenting brain tumors from MRI is a challenging procedure and requires more time. The exact diagnosis of a brain tumor is significant to improve the consequences, but it is challenging to understand the MRI scan. Method: To meet the requirements, a hybrid deep learning network is recommended for segmenting and classifying brain tumors. The main novelty of the research is the proposal of a hybrid network with a residual connection for classifying brain tumors, which can be used to provide proper therapies to affected patients. The recommended model provides accurate classification results, enabling the appropriate clinical decision to enhance the patient';s life expectancy and quality of life. From the public data links, the brain MRI is garnered. Furthermore, the Spatial Attention-based Bilateral Network (SABNet) is proposed for segmenting the tumor region in the input images. Consequently, three sets of features, including morphological, shape, and texture, are extracted, and then feature fusion takes place with the help of optimal weights. Here, the weight is optimally determined by employing the Adaptive Fitness-aided Red Panda Optimization Algorithm (AF-RPOA) for acquiring the weighted feature fusion. At last, the fused features are passed to a Hybrid Residual Network (HRNet), which is built using the Long Short-Term Memory (LSTM) and Capsule Network (CapsNet) architectures. Henceforth, validation of the developed framework is conducted to reveal the simulation outcomes and analysis. Results: According to the evaluation outcome, the proposed AF-RPOA-HRNet achieves an accuracy of 97%. Discussion: The comparative validation of the proposed hybrid deep learning network with conventional works has verified its effectiveness in segmentation and classification. Moreover, the adoption of a weighted feature selection approach has yielded effective performance in classifying brain tumors. Conclusion: On the contrary, the recommended system provides better results that can prove the system';s supremacy in detecting different types of brain tumours.