Mitigating Shortcut Learning in Brain Tumour MRI Classification

Shortcut learning in medical imaging occurs when a classifier exploits acquisition,background or annotation cues that correlate with the label but do not define thepathology. This preprint studies that failure mode in four-class brain tumour magneticresonance imaging (MRI) classification. The model couples a ResNet50V2 feature extractor with a Rotation-Invariant VisionTransformer (RViT) and introduces Pathology-Focused Disentanglement (PFD), whichlearns a soft spatial gate without segmentation masks, and Guided Semantic TokenEvolution (GSTE), which transfers the gate to transformer tokens. The correspondingablations remove PFD and GSTE from each model family. A Secure Hash Algorithm 1(SHA-1) audit removes 297 exact duplicates from 7,023 files, leaving 6,726 imagesacross fixed training, validation and held-out test sets. Held-out test accuracy ranges from 98.52% to 99.22%. Ablation B achieves the highestclassification result, with 99.22% accuracy and 0.9920 macro-F1. Hybrid B achieves98.52% accuracy and 0.9849 macro-F1 and is the only model with tumour-centredGrad-CAM++ evidence in all three reported qualitative cases. All four models remainhighly confident on one shared subtype error. The results expose a discrepancy between classification performance and qualitativespatial evidence. PFD–GSTE alters the observed evidence pathway without improvingthe best held-out classification result. The study evaluates a mitigation-orientedintervention rather than establishing causal removal of shortcut learning; classificationperformance, qualitative saliency and Monte Carlo dropout confidence alone are nottreated as sufficient evidence of shortcut elimination. Code: https://github.com/AnnyaB/HybridResNet50V2-RViT Project website: https://annyab.github.io/pfd-gste-research-article/