Research map / Label-efficient and robust learning / OOD and anomaly detection
Are nuclear masks all you need for improved out-of-domain generalisation? A closer look at cancer classification in histopathology
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes training a ResNet-50 on H&E tiles plus nuclear segmentation masks with an l2 regulariser aligning image and mask embeddings, improving single-domain generalisation for tumour classification across CAMELYON17, BCSS and Ocelot.
Problem
Single-domain generalisation for cancer tumour vs non-tumour tile classification in histopathology, where stain, scanner and hospital differences hurt out-of-domain accuracy.
Approach
Train on H&E images together with binary nuclear masks (and 50% mask-times-image inputs), adding an l2 loss that pulls the pre-GAP embedding of the image toward that of its nuclear mask, so the model relies on nuclear shape and organisation rather than colour/texture; no masks are needed at inference.
Key results
- On CAMELYON17 with photometric augmentation, Ours-Aug averages 91.8% out-of-domain accuracy vs 89.1% for L2D-Aug and 85.5% for ERM-Aug.
- On BCSS, Ours-Aug averages 78.8% vs 75.1% for L2D-Aug (p=4e-5 paired t-test).
- On Ocelot, Ours-Aug averages 70.6% vs 69.1% for L2D-Aug (p=0.044); the gap is smaller and the method underperforms L2D on Endometrium and Kidney.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with nuclear masks and ℓ₂ feature alignment, with photometric augmentation | ResNet-50 · CAMELYON17 OOD | accuracy: 89.1±1.1 -> 91.8±0.9 | Helped |
| Train with nuclear masks and ℓ₂ feature alignment, with photometric augmentation | ResNet-50 · BCSS OOD | accuracy: 75.1±1.3 -> 78.8±1.6 | Helped |
| Train with nuclear masks and ℓ₂ feature alignment, with photometric augmentation | ResNet-50 · Ocelot OOD | accuracy: 69.1±1.1 -> 70.6±1.0 | Helped |
| Add ℓ₂ feature-alignment regularisation (ablation) | Ours with nuclear masks and photometric augmentation · CAMELYON17 OOD | accuracy: 87.8±1.7 -> 91.8±0.9 | Helped |
| Add ℓ₂ regularisation and mask-times-input augmentation (ablation) | Ours-MO-Aug · CAMELYON17 OOD | accuracy: 84.5±2.7 -> 91.8±0.9 | Helped |
Surprising findings
- Replacing intranuclear texture with a nuclear-like colour barely changes accuracy, while whitening nuclei reduces it, suggesting white-background artefacts matter.
- Subtracting mask embeddings at test time drops the proposed method's accuracy below random guessing on CAMELYON17, showing its features are close to mask features.
Limitations
- Evaluated only on one classification task (tumour vs non-tumour) and one cancer type family, lymph node and breast tissue.
- Full experiments only with ResNet-50; ViT-Tiny preliminary results are mixed and used untuned hyperparameters.
- Underperforms L2D on Endometrium and Kidney in Ocelot.
Experimental setup
Datasets: CAMELYON17 BCSS Ocelot CoNSeP
Benchmarks: CAMELYON17 BCSS Ocelot
Metrics: tile-level accuracy accuracy drop under image corruptions PGD adversarial accuracy recall
Baselines: ERM Macenko normalisation HoVer-Net pretrained ResNet-50 RandStainNA RSC L2D DDCA
Builds on: RandStainNA Learning to Diversify HoVer-Net Representation Self-Challenging
Closest papers by meaning
- Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.86)
- Incremental Nuclei Segmentation from Histopathological Images via Future-class Awareness and Compatibility-inspired Distillation (CVPR 2024 · similarity 0.85)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.84)
- UniCell: Universal Cell Nucleus Classification via Prompt Learning (AAAI 2024 · similarity 0.84)
- Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning (AAAI 2025 · similarity 0.84)
- Rotation-Agnostic Image Representation Learning for Digital Pathology (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.