Healthy Counterfactual Generation via Diffusion Inpainting for Mammography Classification

1 Introduction

• False negatives are associated with delayed diagnosis and worse outcomes, while CAD predictions may rely on spurious correlations rather than lesion-specific cues.For clinical deployment, the introduction says systems should improve sensitivity and base predictions on regions of interest.Source: §1 · pp. 1–2
• The study addresses the underexplored use of healthy counterfactuals for augmentation by replacing annotated lesion regions while preserving surrounding anatomy.The method generates plausible normal counterparts of abnormal images rather than synthesizing lesions.Source: §1 · p. 2
• The approach is evaluated across four classifier architectures with distinct inductive biases: ConvNeXt, ViT, Mammo-CLIP, and FPN-MIL.The selected models span convolutional, transformer, mammography-specific image-encoder, and multiple-instance-learning approaches.Source: §1 · p. 2
• Sensitivity gains over standard augmentation baselines are consistent, particularly at 80% specificity on VinDr-Mammo.The passage presents this operating point as clinically relevant and frames the strategy as supporting fewer missed detections.Source: §1 · p. 2

2 Method

• The three-stage pipeline trains on healthy mammograms, inpaints lesion regions using bounding-box-derived masks, and incorporates generated counterfactuals into classifier training.The pipeline overview identifies the DDPM as the healthy-tissue model and RePaint as the mask-guided inpainting step.Source: §2 · p. 2, Figure 1 · p. 3

2.1 Denoising Diffusion Probabilistic Model (DDPM)

• A DDPM with a U-Net learns normal breast anatomy by predicting injected Gaussian noise during training and iteratively denoising random noise at inference.Training uses BI-RADS 1 mammograms exclusively, with an MSE noise-prediction objective.Source: §2.1 · p. 3

2.2 RePaint Inpainting

• RePaint re-noises known pixels at each step and updates masked regions through the learned reverse transition, combining preserved tissue with newly generated content.The element-wise update uses the mask to combine the preserved region and the newly generated image content.Source: §2.2 · p. 3

3.1 Dataset

• VinDr-Mammo contributes 5,000 exams and 20,000 mammograms, split into 16,000 training and 4,000 test images.Each exam contains four standard CC and MLO views; images were resized to 512 × 512 for generation and later restored to 1520 × 912.Source: §3.1 · pp. 3–4

3.2 Counterfactual Generation

• The DDPM was trained on BI-RADS 1 images, and RePaint used bounding-box-derived binary masks to replace annotated lesion regions while preserving surrounding tissue.One counterfactual was generated per anomalous annotated mammogram; sampling used 500 respaced timesteps.Source: DDPM Training · p. 4, Sampling using RePaint · p. 4

3.3 Counterfactual Augmentation

• The evaluation spans ConvNeXt, ViT, Mammo-CLIP, and FPN-MIL, selected to represent different classifier inductive biases.ConvNeXt and ViT used ImageNet-pretrained weights; Mammo-CLIP and FPN-MIL used VinDr-Mammo-pretrained EfficientNet backbones.Source: §3.3 · p. 4
• Adding one healthy counterfactual per annotated image supplied 1,410 negative samples, increasing the negative class by 13%.Positive-class weights were recalculated after augmentation to reflect the changed class distribution.Source: §3.3 · p. 5

4.1 Healthy Counterfactual Generation Quality Assessment

• Distributional FID and FRD results suggested edits remained close to real healthy mammograms, and patch scores were reported as better than a study using different datasets and patch types.The comparison with Osuala et al. was described as indicating competitive realism, despite protocol differences.Source: Distributional Metrics · p. 5, Table 1 · p. 6
• The blind reader study compared 70 counterfactuals with 30 real healthy mammograms, while six radiologists assessed a balanced subset of 100 synthetic images.Three examples received unanimous 5/5 realism ratings from all reviewing radiologists.Source: Expert Reader Study · p. 6, Figure 3 · p. 7
• Real and counterfactual mammograms received broadly aligned ratings, although counterfactuals shifted slightly lower for full-image realism and bounding-box healthiness.Both bounding-box rating distributions remained concentrated in the 3–4 range.Source: Expert Reader Study · p. 6, Figure 2 · p. 7

4 Results

• The reader noted source-dataset defects, and the authors did not perform matched-pair prediction-consistency analysis because its interpretation was limited in this setting.Reported source issues included low contrast, unrealistic soft-tissue appearance, implausible axillary superposition, and incomplete breast coverage.Source: §4 · p. 5, Expert Reader Study · p. 6

4.1 Healthy Counterfactual Generation Quality Assessment

• Density D cases received the strongest overall realism ratings, while larger masked lesions tended to receive lower ratings.Density C cases had the largest proportion of upper-end ratings compared with densities A and B.Source: Expert Reader Study · pp. 6–7, Figure 4 · p. 8

4.2 Data Augmentation

• Recall improved across all four classifiers, with gains of 4.6% for ConvNeXt and 1.9% for FPN-MIL; recall at 80% specificity rose by 0.6%–1.3%.Balanced accuracy and F1-score also improved modestly across all models.Source: Classification Performance · p. 7, Table 2 · p. 8
• ROC-AUC improved modestly for ViT and Mammo-CLIP but remained nearly unchanged for ConvNeXt and FPN-MIL; specificity decreased for three models.Specificity improved slightly for ViT, while sensitivity gains were consistent across models.Source: Classification Performance · pp. 7–8

• The framework generated lesion-free counterfactuals by training a DDPM on healthy mammograms and RePaint-inpainting annotated lesions while preserving patient-specific anatomy.Source: §5 · p. 8
• Distributional metrics and expert reader assessments found generated images consistent with real healthy mammograms.Source: §5 · p. 8
• Sensitivity improved consistently across diverse classifiers, especially at 80% specificity; F1-score, balanced accuracy, and ROC-AUC gains were more modest, while some models lost specificity.Source: §5 · p. 9