HYBRID DEEP LEARNING FRAMEWORK FOR ACCURATE MULTI CLASS BRAIN TUMOR MRI CLASSIFICATION AND PRECISE SEGMENTATION WITH OPTIMIZED FEATURE EXTRACTION

Abstract

Brain tumor detection using MRI images plays a pivotal role in the early diagnosis and treatment planning. The conventional deep learning models, despite
it shows a promise often struggles to capture the complex, multi scale characteristics of the tumor regions. We proposed a hybrid pipeline that combines the Fast Haar
Curvelet Transform (FHCT), Deep Neural Networks (DNN), and the DenseNet based architectures. Preprocessing involved histogram equalization and Gaussian
filtering to normalize the intensity variations. FHCT was applied to extract the multi scale, multidirectional features that specifically capture the curved and irregular
tumor boundaries. These features were refined using a deep DNN to enhance the nonlinear separability and reduce redundancy. Principal Component Analysis (PCA)
further has reduced the dimensionality, where it improves the computational efficiency. AdapDenseNet was then employed for multiclass tumor classification (normal,
benign, malignant), while a U-Net architecture has enabled the pixel level segmentation. The proposed method shows superior performance in both the classification
and segmentation tasks. Experimental evaluations showed a classification accuracy of 0.95, precision of 0.94, recall of 0.93, F1-score of 0.94, and the Dice coefficient
of 0.95, which performs better than the existing CNN, U-Net, and the Curvelet-DNN approaches. Tumor boundaries were delineated with higher precision, and the
computational time was reduced due to the efficient feature refinement.