Data-Driven Brain Tumor Classification Using CNNs with Trainable Gabor and LoG Feature Mining Layers

Abstract

Magnetic resonance imaging (MRI) is an important method for reproducible brain tumour diagnosis. Conventional convolutional neural networks (CNNs) and transfer learning models have been demonstrated to yield high accuracy, but are not yet without certain shortcomings in terms of domain shifts, interpretability, and suboptimal feature generalization at a lower level. This work introduces a novel hybrid architecture using trainable CNN coupled with Gabor and Laplacian of Gaussian (LoG) modules to sequentially learn textural and edge-based features. The model is evaluated against an open-source Brain Tumor MRI dataset with conditions including glioma, meningioma, pituitary and no-tumor. The proposed model achieves an overall accuracy of 96%, with better accuracy compared to baseline CNN performance and high interpretability in terms of learned kernels and saliency maps which provide clinically interpretably outputs.