Explainable AI Based Dual-Domain Deep Network for Brain Tumor Classification

The principal contributions of this work are summarized as follows:
• A dual-domain fusion architecture that couples a frozen EfficientNetB0 spatial encoder with a purpose-built wavelet-domain CNN, addressing the exclusively spatial domain limitation shared by recent approaches.

• An explicit four-subband wavelet feature-engineering pipeline in which the LL,LH,HL,HH decomposition is presented as a four-channel tensor, enabling the network to learn inter-subband correlations — such as the co-occurrence of horizontal and diagonal detail energy that characterizes a spiculated infiltrative margin—that a spatial-only CAM-interpretable model cannot access directly.

• A dual-modality explainability protocol integrating Grad-CAM and a dual-input-faithful adaptation of LIME, extending the CAM-only interpretability practice, independent attribution method and explicit input-space faithfulness for the frequency branch