CNS Tumor Classification Using Hierarchical ResNet and EfficientNet Cascade Networks

Accurate subtype classification of central nervoussystem (CNS) neoplasms is critical for guiding treatment decisions,yet remains challenged by morphological overlap across WHO-defined grades, inter-observer variability in neuropathologicalinterpretation, and the inherent biological continuum that linkslow-grade and high-grade gliomas. We present a hierarchicaltwo-stage deep learning framework—combining ResNet-18 andEfficientNet-B0 backbones—for 19-class CNS tumor classificationon 1,999 de-identified institutional samples spanning the full WHO5th Edition (CNS5) taxonomy.Stage 1 (binary triage) achieves 98.6% accuracy for ResNet-18 and 99.0% for EfficientNet-B0, with ResNet-18 yielding zeromissed-tumor false negatives (FN = 0), a clinically critical safetyproperty. Stage 2 (19-class subtype) achieves 89.4% accuracy withMacro-F1 = 0.846 for ResNet-18 and 90.4% / 0.865 for EfficientNet-B0. Both models attain perfect accuracy on brain_tumor_misc(200/200), other_lesions (46/46 ResNet-18), pituitary (222/223),schwannoma (172/173), and glioma (201–202/205), while exhibitingsystematic confusion at two clinically significant boundaries: the as-trocytoma/glioblastoma interface (31.0% of astrocytoma cases mis-graded as GBM by ResNet-18) and the oligodendroglioma/GBMinterface (31.7% mis-classification rate for ResNet-18).We provide a biologically grounded analysis of these failuremodes, linking them to WHO molecular criteria and archi-tectural limitations of fixed-receptive-field CNNs. We furthersurvey contemporary mitigation strategies—including squeeze-and-excitation (SE) networks, convolutional block attentionmodules (CBAM), Swin Transformer hybrids, Vision Mamba, andMonte Carlo dropout uncertainty quantification—and evaluatetheir suitability for each observed failure mode. Our two-thresholdrisk-stratification framework transforms model outputs intoclinically actionable GBM probability scores, enabling integrationinto routine neuropathology workflows while preserving humanoversight at biological continuums.Index Terms—CNS tumors, ResNet-18, EfficientNet-B0, hier-archical classification, WHO CNS5, glioma grading, attentionmechanisms, Swin Transformer, Vision Mamba, uncertaintyquantification, Monte Carlo dropout, diagnostic boundaries, deeplearning limitations.