Visualizations - Mitigating Shortcut Learning in Brain Tumour MRI Classification

← Back to article

Visual assets from

Mitigating Shortcut Learning in Brain Tumour MRI Classification

4 Charts

4 Images

9 Tables

Evidence-aware medical AI · PFD–GSTE Select a stage to inspect the intervention

Fixed four-class MRI benchmark

Begin from a leakage-aware split created after exact duplicate removal; preserve Kaggle Testing as the held-out test set.

6,726 unique1,284 test4 classes

#1 Table

Study fact	Verified value
Raw images discovered	7,023
Exact duplicates removed	297
Unique images retained	6,726
Held-out test images	1,284
Best test accuracy	99.22%
Best macro-F1	0.9920
Fewest test errors	10

#2 Table

Model	CNN pathway	Transformer pathway	Guidance intervention
Hybrid A	ResNet50V2 descriptor	49 CNN feature tokens	PFD mask guides transformer-facing tokens
Ablation A	Matched A-family backbone	Same tokenisation	PFD–GSTE removed
Hybrid B	PFD-gated ResNet50V2 descriptor	196 raw-image patch tokens before optional pooling	PFD gates CNN evidence and guides patch tokens
Ablation B	Matched B-family backbone	Same tokenisation	PFD–GSTE removed

#3 Stack

Hybrid A architecture showing ResNet50V2, PFD guidance, 49 CNN feature tokens, RViT processing, and late fusion.

Hybrid B architecture showing PFD-gated CNN evidence, 196 raw-image patch tokens, GSTE guidance, optional dynamic weighted pooling, RViT processing, and late fusion.

#4 Table

Property	Hybrid A	Hybrid B
Initial transformer tokens	49 CNN feature tokens	196 raw-image patch tokens
CNN descriptor	Ungated	PFD-gated
Guidance reach	Transformer-facing pathway	CNN and transformer pathways
Token count	Fixed	Optional concentration-aware pooling
Trainable parameters	26.68M	26.58M

#5 Table

Split	Glioma	Meningioma	Pituitary	No tumour	Total
Train	1,057	1,064	1,152	1,080	4,353
Validation	264	267	288	270	1,089
Test	299	304	300	381	1,284

#6 Stack

Original Kaggle split composition.

Original Kaggle training and testing class counts before duplicate removal. — Original Kaggle split composition.

Class distribution after exact duplicate removal. — Class balance after deduplication.

#7 Image

Representative brain MRI images inspected during the rule-based raw-image audit.

Representative raw-image audit panel. The report recorded zero unreadable, too-dark, too-bright, or low-contrast files under its rule-based checks.

#8 Table

Component	Setting
Optimiser	AdamW
CNN learning rate	$10^{-4}$
Transformer and fusion learning rate	$5\times10^{-4}$
Weight decay	0.01
Objective	Cross-entropy with label smoothing 0.05
Schedule	Cosine annealing, $\eta_{\min}=10^{-6}$
Batch size	32
Maximum epochs	100
CNN warm-up	Frozen for first 5 epochs
Early stopping	Validation macro-F1, patience 10
Gradient clipping	1.0
Random seed	42

#9 Stack

Guided-model learning curves. Hybrid B reached its best validation macro-F1 at epoch 14.

Training and validation loss, accuracy, and macro-F1 curves for Hybrid A and Hybrid B. — Guided-model learning curves. Hybrid B reached its best validation macro-F1 at epoch 14.

Training and validation loss, accuracy, and macro-F1 curves for Ablation A and Ablation B. — Ablation learning curves. Late fluctuations support validation-based early stopping rather than final-epoch selection.

#10 Chart

Held-out performance explorer

Switch metrics to compare all four variants. Percentage axes are intentionally zoomed and labelled; the error view uses the absolute count from 1,284 test images.

View in article →

#11 Table

Model	Params	Accuracy	Macro-F1	Loss	Specificity	$\kappa$	MCC	Errors	Best epoch
Hybrid A	26.68M	0.9875	0.9875	0.0803	0.9959	0.9833	0.9833	16	43
Hybrid B	26.58M	0.9852	0.9849	0.0753	0.9952	0.9802	0.9803	19	14
Ablation A	26.68M	0.9875	0.9873	0.0847	0.9959	0.9833	0.9834	16	30
Ablation B	26.58M	0.9922	0.9920	0.0668	0.9975	0.9896	0.9896	10	42

#12 Chart

Classification–localisation trade-off

Horizontal position is held-out test accuracy. Vertical position is the number of Grad-CAM++ judgements marked “Correct” in the three selected report cases. The vertical axis is qualitative and n=3—not a segmentation score.

View in article →

#13 Table

Case	Hybrid A	Hybrid B	Ablation A	Ablation B
Correct meningioma	GC++ wrong; rollout partially correct	GC++ correct; rollout wrong	GC++ wrong; rollout partially correct	GC++ correct; rollout wrong
Challenging glioma	GC++ wrong; rollout wrong	GC++ correct; rollout partially wrong	GC++ correct; rollout partially wrong	GC++ partially wrong; rollout partially wrong
Pituitary → meningioma error	GC++ wrong; rollout wrong	GC++ correct; rollout partially correct	GC++ wrong; rollout partially correct	GC++ wrong; rollout partially correct

#14 Chart

Confidence on the shared pituitary → meningioma error

Mean confidence over 20 MC Dropout passes. The axis is zoomed to 90–100% and must not be interpreted as a calibration curve.

View in article →

#15 Table

Dimension	Verified now	Completion gate
Research integrity	Exact reported values, negative ablation outcomes, authorship, and limitations are preserved	Keep all claims traceable to report, code, or released artefacts
Open artefacts	Public code, checkpoints, model cards, PyPI package, report, and machine-readable article data	Publish this article repository with a versioned release and build check
Statistical robustness	One seeded training run per variant	Run at least five matched seeds per variant; report mean, standard deviation, 95% confidence intervals, and paired within-family deltas
Fair comparison	Guided variants have same-family ablations	Use the same loader policy, preprocessing manifest, checkpoint rule, and deterministic settings in the repeated-seed rerun
Calibration	Twenty-pass MC Dropout summaries for selected examples	Archive validation and test probabilities; report NLL, multiclass Brier score, ECE, reliability diagrams, and validation-fitted temperature scaling
External generalisation	A five-image external meningioma sample exists only as a qualitative demo	Evaluate a provenance-audited external cohort with compatible labels and no overlap with the development benchmark
XAI validation	Grad-CAM++ and Attention Rollout are reviewed on three selected cases	Evaluate a preregistered masked subset with Dice, IoU, pointing-game or localisation-error metrics plus saliency sanity checks
Clinical evidence	None claimed	Independent clinical governance, representative cohorts, and prospective validation would be required before any clinical claim

#16 Table

Resource	Access
Model and experiment code	GitHub repository
Browser demonstration	Hugging Face Space
Reusable modules	PFD–GSTE on PyPI
Machine-readable website data	`app/src/content/assets/data/` in this article repository
Evidence-upgrade protocol	<code>VALIDATION_PLAN.md</code> in the article repository