PruDensNet: a parameter efficient depthwise separable CNN for MRI-based brain tumor classification

This study introduces PruDensNet, a parameter-efficient, depthwise-separable convolutional network for magnetic resonance imaging (MRI)-based brain-tumor classification, motivated by compute and latency-sensitive deployments where efficiency can improve throughput and cost. The architecture integrates lightweight channel and spatial attention with Gaussian Error Linear Unit (GELU) activations and a compact global average pooling (GAP)-Dense head, totaling approximately 1.46 M parameters. Architectural effects are isolated from capacity by equalizing parameter budgets across baselines through a no-operation padding mechanism. A reproducible curation pipeline standardizes labels, removes near-duplicates, and applies a stratified split. Training follows a curriculum-regularized recipe combining MixUp, CutMix, random erasing, CLAHE, and label smoothing, with AdamW optimization, warmup-cosine decay, gradient clipping, mixed precision, and test-time augmentation. On a four-class Brain Tumor MRI benchmark (glioma, meningioma, pituitary, no tumor), PruDensNet attains test accuracy 96.05% and validation accuracy 97.27% with competitive per-class metrics, outperforming matched-capacity Convolutional Neural Network (CNN) and Transformer baselines. The results indicate a favorable accuracy-footprint trade-off that can support cost- and latency-sensitive clinical workflows, subject to external validation and hardware-specific benchmarking.