An explainable modified convolutional mixer neural network-based deep learning framework for accurate brain tumor detection and classification.

Authors

Abstract

Brain tumors (BTs) pose a severe health risk due to their heterogeneous nature and low survival rates. Traditional models often struggle with computationally expensive and require extensive training time. To address this limitation, this paper introduces a novel Explainable Modified Convolutional Mixer Neural Network (EM-ConvMixer+Net) for BT detection and classification. A total of 10,087 MRI images are collected from two publicly available datasets: Brain Tumor MRI Dataset (7,023 images) and Figshare Brain Tumor Classification Dataset (3,064 images). Denoising, contrast stretching, labeling, and augmentation are the preprocessing steps utilized to enhance image quality. Tumor segmentation is performed using the Adopted Region Growing method, which ensures accurate extraction of tumor features. The convolutional operations mixer block (ConvMixer+) removes redundant features and enhances rich features for improved classification efficiency. Cross Attention Network (CANet) fuses multi-scale encoder features, which capture long-range dependencies for improved feature representation. External Attention Network (EANet) refines these spatial features by emphasizing relevant regions and reducing noise for better generalization. Finally, Classification is handled by a Modified Artificial Neural Network (M-ANN), while Gradient-weighted Class Activation Mapping (Grad-CAM) is utilized for explainability to enhance model interpretability. The experimental validation is conducted for the identification of the precise outcome of the proposed model with an accuracy of 98.86% and F1-score of 98.44%, as demonstrated through extensive comparative and ablation studies. Overall, the EM-ConvMixer+Net technique presents a reliable, interpretable, and computationally efficient framework for advancing clinical brain tumor diagnosis.

Tags

Topics