Diffusion models generate healthy mammograms to improve breast cancer detection
Healthy Counterfactual Generation via Diffusion Inpainting for Mammography Classification
Computer Vision and Pattern Recognition
Summary
False negatives in breast cancer screening happen when cancer is missed, delaying treatment. To help, the authors developed a technique that digitally removes lesions from mammogram images to create realistic healthy versions. These new images help train AI systems to better detect cancer by showing more examples of what healthy breasts look like. Testing on various AI models showed improved ability to spot cancer, which could lead to more reliable breast cancer detection tools.
What this means in practice
- •For medical imaging teams: Improve breast cancer detection tools by augmenting training data with realistic healthy mammograms generated from lesion removal.
- •For radiology software developers: Enhance performance of CAD systems by integrating diffusion-based counterfactual images to reduce missed cancer cases in mammogram analysis.
Tested on one dataset.
Authors
Inês Cruchinho Garcia, Mariana Mourão, Francisco Maria Calisto, Carlos Santiago, Jacinto Nascimento
Abstract
False negatives remain a critical limitation of computer-aided diagnosis (CAD) systems for breast cancer screening due to delayed detection and treatment. To address this issue, we propose a counterfactual data augmentation strategy that generates healthy mammograms by "erasing" lesions from anomalous images, thereby enriching the training distribution. We train a Denoising Diffusion Probabilistic Model on BI-RADS 1 (healthy) mammograms and use a RePaint-based sampling strategy to inpaint realistic normal tissue within annotated lesion bounding boxes. The resulting healthy counterfactuals replace annotated lesion regions with realistic healthy tissue while preserving patient-specific anatomical structure, as supported by similarity metrics between real and generated images. Image realism was further assessed by radiologists and found to be consistent with the original dataset quality. We evaluate counterfactual augmentation across four representative classifier architectures: a convolutional neural network (ConvNeXt), a vision transformer (ViT), a vision-language model pre-trained on mammogram-report pairs (Mammo-CLIP) and a multi-scale attention-based multiple-instance learning framework (FPN-MIL). Experiments conducted on the VinDr-Mammo dataset show improvements in sensitivity across all architectures, particularly at 80\% fixed specificity, contributing towards more reliable CAD systems for breast cancer. Code is available at: https://github.com/ines03garcia/diffusion-based-counterfactual-generation.