AdaptiveResCNN: Preserving Multi-Frequency Features and Adaptive Scaling for Robust Brain Tumor MRI Analysis

Brain tumors exhibit high variability in size, shape, texture, and frequency-specific patterns, complicating accurate classification from MRI. This study introduces AdaptiveResCNN, a lightweight residual CNN that addresses these challenges through explicit frequency-domain decomposition and input-adaptive multi-scale processing, without data augmentation or pre-training. Conventional pooling is replaced by DecompResPool, which applies Haar-wavelet decomposition to preserve LL, LH, HL, and HH subbands during downsampling, retaining both anatomical context and tumor edges/textures. Complementing this, the MultiPathDilBlock employs channel-grouped parallel dilated depthwise convolutions (rates 1, 2, 3) fused dynamically by the content-adaptive ScaleDynamicFuse module with Squeeze-and-Excitation attention. The architecture delivers robust binary and multi-class tumor identification and typing in clinical MRI analysis.