Research map / Adversarial robustness and security / Certified and empirical robustness
NaFV-Net: An Adversarial Four-view Network for Mammogram Classification
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 NaFV-Net, a four-view mammogram classifier that uses contralateral (unaffected-side) images to build adversarial samples via lesion-region mixing, combined with a mass segmentation module and a hybrid CNN-Transformer multi-view classifier on DINOv2 features.
Problem
Breast cancer benign/malignant classification from contrast-enhanced mammography (CEM) images, where prior multi-view methods mostly use only affected-side CC and MLO views and ignore the unaffected side that radiologists examine.
Approach
Uses real images from the non-affected side, with the affected-side mass pasted at the symmetric location via soft masks, as adversarial samples; trains with a focal plus KL-based robustness loss, and classifies with a hybrid CNN-Transformer that cross-attends across CC and MLO views on both sides.
Key results
- On the internal Tongji dataset, NaFV-Net reaches AUC 0.935, precision 0.902 and recall 0.899 under 10-fold cross-validation, above the best baseline DCHA-Net (AUC 0.907).
- On the public CDD-CESM dataset, NaFV-Net reports AUC 0.883 versus 0.871 for DCHA-Net.
- For mass segmentation on the internal dataset, the method reaches Dice 88.18% and mIoU 79.16%, above SwinUNet (85.14 / 76.31).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the asymmetric sensitivity module (ablation) | NaFV-Net segmentation and classification · Internal dataset | Dice: 4.0% | Helped |
| add the asymmetric sensitivity module (ablation) | NaFV-Net segmentation and classification · Internal dataset | AUC: 1.2% | Helped |
| add the attention module (ablation) | NaFV-Net segmentation and classification · Internal dataset | Dice: 9.1% | Helped |
| add the attention module (ablation) | NaFV-Net segmentation and classification · Internal dataset | AUC: 1.6% | Helped |
| use affected- and unaffected-side four-view images instead of affected-side CC and MLO only (ablation) | NaFV-Net classification · Internal dataset | AUC: 5.3% | Helped |
| perturb the position masks of unaffected-side images instead of keeping them symmetrically aligned (ablation) | NaFV-Net classification · Internal dataset | AUC: 2.0% | Helped |
| fine-tune DINOv2 on the breast mammography datasets (ablation) | NaFV-Net · Internal dataset | AUC: 91.47 -> 93.52 | Helped |
Surprising findings
- Using only affected-side views yields a 5.3% drop in classification performance versus the four-view model, but the paper does not report a benchmark-level comparison beyond its own ablations.
Limitations
- Evaluated on a single-center internal dataset of 479 patients; the authors note plans for multi-center validation.
- Mass segmentation could only be tested on the internal dataset because the public dataset lacks mass annotations.
- Hyperparameters and baseline training settings were aligned to the proposed model rather than tuned per baseline.
Experimental setup
Datasets: Tongji internal CEM dataset CDD-CESM CBIS-DDSM INbreast CMMD
Benchmarks: Tongji internal dataset CDD-CESM
Metrics: AUC Precision Recall Dice mIoU
Baselines: Deep MIL GMIC Sun et al. 2022 PHBreast BRAIxMVCCL DCHA-Net DINOv2 UNet UNet++ TransUNet UNeXt SwinUNet
Builds on: DINOv2 Focal loss MedSAM GMIC
Closest papers by meaning
- Prompt2Perturb (P2P): Text-Guided Diffusion-Based Adversarial Attack on Breast Ultrasound Images (CVPR 2025 · similarity 0.84)
- Attacking CNNs in Histopathology with SNAP: Sporadic and Naturalistic Adversarial Patches (Student Abstract) (AAAI 2024 · similarity 0.83)
- Learning Low-Rank Feature for Thorax Disease Classification (NeurIPS 2024 · similarity 0.83)
- Adversarial Learning Under Hybrid Perturbations for Robust Acute Lymphoblastic Leukemia Classification (AAAI 2025 · similarity 0.82)
- The Effect of Intrinsic Dataset Properties on Generalization: Unraveling Learning Differences Between Natural and Medical Images (ICLR 2024 · similarity 0.82)
- Curriculum Fine-tuning of Vision Foundation Model for Medical Image Classification Under Label Noise (NeurIPS 2024 · similarity 0.82)
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.