AN IMPROVED XCEPTION BASED FEATURE LEARNING AND HONEYBADGER OPTIMIZED LONG SHORT-TERM MEMORY FOR ACCURATE BRAIN TUMOR CLASSIFICATION WITH DATASET BALANCING APPROACH

A brain tumour (BT) is a severe and lethal disease that significantly reduces human lifespan. Magnetic resonance imaging (MRI) is a commonly employed imaging technique for the early detection of tumours. The segmentation and classification of brain tumours (BTs) through manual and traditional methods is a labour-intensive and subjective process, necessitating the involvement of expert radiologists for evaluation, which may result in prediction inaccuracies. Class imbalance presents a notable challenge in MRI datasets, impacting the efficacy of the classification system. This paper presents an optimised deep learning approach, referred to as Honey Badger optimised Long Short-Term Memory (HBLSTM), aimed at the detection of BT through effective segmentation and feature extraction methodologies. The preprocessing steps are conducted on the collected CE-MRI dataset to reduce noise through Gaussian filtering. The class imbalance issue is addressed through the application of the Adaptive Synthetic Sampling (ADASYN) model. The system employs the Spatial and Channel attention-based Three-Dimensional U-shaped Network (SC3DUNet) for tumour segmentation. The segmentation images utilise the most discriminative features through the Spatial Pyramid Pooling centred Xception Network (SPPXNet). The essential features are subsequently selected utilising the Diagonal Linear Uniform and Tangent Flight-based Butterfly Optimisation Algorithm (DTBOA). The tumour classes are classified utilising the HBLSTM algorithm. The experimental results demonstrate the efficacy of our hybrid deep learning models, achieving an average accuracy of 99.81% in tumour detection, surpassing current state-of-the-art models.