Source-linked AI summary
Conditional Infilling GANs for Data Augmentation in Mammogram Classification
Eric Wu, Kevin Wu, David Cox, William Lotter
TL;DR
Limited mammography datasets and class imbalance constrain breast cancer detection research. This paper uses a class-conditional infilling GAN to synthesize lesions for mammogram augmentation, and reports higher AUC than baseline and traditional augmentation alone.
Problem
Mammography datasets are limited by privacy restrictions and annotation costs, while normal images substantially outnumber images with findings.
Source: §1 · p. 1Method
The authors use a multiscale class-conditional infilling GAN to generate lesions on mammogram patches and augment classifier training data.
Source: §2 · p. 4, §3.2 · p. 6Results
GAN augmentation combined with traditional augmentation reached AUC 0.896, exceeding the baseline by 0.014 AUC and traditional augmentation by 0.009 AUC.
Source: §3.3 · pp. 6–7Takeaways & Limitations
The authors report that ciGAN generates realistic lesions and improves subsequent classification performance above traditional augmentation techniques.
Source: §4 · p. 7
Abstract
from arXiv · showhide
Deep learning approaches to breast cancer detection in mammograms have recently shown promising results. However, such models are constrained by the limited size of publicly available mammography datasets, in large part due to privacy concerns and the high cost of generating expert annotations. Limited dataset size is further exacerbated by substantial class imbalance since "normal" images dramatically outnumber those with findings. Given the rapid progress of generative models in synthesizing realistic images, and the known effectiveness of simple data augmentation techniques (e.g. horizontal flipping), we ask if it is possible to synthetically augment mammogram datasets using generative adversarial networks (GANs). We train a class-conditional GAN to perform contextual in-filling, which we then use to synthesize lesions onto healthy screening mammograms. First, we show that GANs are capable of generating high-resolution synthetic mammogram patches. Next, we experimentally evaluate using the augmented dataset to improve breast cancer classification performance. We observe that a ResNet-50 classifier trained with GAN-augmented training data produces a higher AUROC compared to the same model trained only on traditionally augmented data, demonstrating the potential of our approach.
1 Introduction
The introduction frames scarce, imbalanced mammography datasets as a challenge for cancer classification and proposes GAN-based lesion synthesis as an augmentation strategy. It reports higher AUC with GAN-augmented training data than with the baseline and traditional augmentation alone.
1 Introduction
- Medical imaging datasets can be difficult to collect because of privacy restrictions, expert annotation needs, and data distributed across sites; class imbalance compounds the challenge.The passage notes that normal images dramatically outnumber images with findings.Source: §1 · p. 1
- Figure 1 illustrates both malignant-to-non-malignant and non-malignant-to-malignant patch synthesis, with contextual information smoothing segmentation-mask borders.The malignant insertions use randomly selected segmentations from other malignant patches.Source: Figure 1 · p. 2
- Because many tumors are localized in seemingly normal tissue, the proposed augmentation samples a location in a normal mammogram and synthesizes a lesion there.Generating a smaller patch rather than an entire image is also presented as a benefit for the generative model.Source: §1 · p. 3
- GAN-based augmentation improved mammogram patch classification by 0.014 AUC over the baseline and 0.009 AUC over traditional augmentation alone.The approach also generated synthetic patches at 256×256 pixels.Source: §1 · p. 3
2 Proposed Approach: Conditional Infilling GAN
The proposed ciGAN infills a segmented lesion region in a mammogram patch, conditioned on the desired class. Its multiscale design and combined losses target stable high-resolution synthesis with smooth lesion-context boundaries.
2 Proposed Approach: Conditional Infilling GAN
- ciGAN uses multiscale feature generation, infills only the segmented lesion region, and shares learned features between malignant and non-malignant cases through class conditioning.The multiscale generator is based on a cascading refinement network.Source: §2 · p. 4
2.1 Architecture
- The generator combines a corrupted context image, lesion mask, and two class-label channels, progressively refining its representation to 256×256 pixels.The discriminator uses an inverse convolutional structure to classify 256×256 images.Source: §2.1 · p. 4
2.2 Training Details
- Training uses 256×256 lesion-containing patches to focus evaluation on GAN-augmented data and build on patch-level pretraining's reported generalization benefits.Source: §2.2 · p. 4
- The training objective combines VGG-19 feature-map matching, adversarial loss, and a boundary loss that encourages smooth transitions between infilled regions and surrounding context.Feature loss compares maps at pool1, pool2, and pool3; boundary loss measures L1 difference at the boundary.Source: §2.2 · p. 5
- The generator is first pretrained on feature loss for 10,000 iterations, then trained with the discriminator on all losses for another 100,000 iterations.The loss coefficients are 1.0 for GAN loss, 10.0 for feature loss, and 10000.0 for boundary loss.Source: §2.2 · p. 5
3 Experiments
The experiments use DDSM mammogram patches to evaluate ciGAN-based augmentation for cancer detection. The combined ciGAN and traditional augmentation regime achieves an AUC of 0.896, outperforming both comparison models.
3.1 DDSM Dataset
- DDSM contains 10,480 images, including 1,832 malignant and 8,648 non-malignant cases, with patches labeled using dataset segmentation masks.Non-malignant patches include benign and non-lesion patches; the dataset is split 80%/10%/10%, and sampled 256x256 patches must contain more than 75% breast tissue.Source: §3.1 · pp. 5–6
3.2 GAN-based data augmentation
- The ciGAN regime doubles the dataset by inserting malignant lesions into non-malignant patches and replacing lesions in malignant patches with non-malignant content.Training begins with equal proportions of real and synthetic data, then increases the real-data proportion every 1,000 iterations until it reaches 90%; the authors report this helps prevent early overfitting and supports later generalization.Source: §3.2 · p. 6
3.3 Results
- AUC 0.896 with ciGAN plus traditional augmentation exceeds the no-augmentation baseline by 0.014 AUC and traditional augmentation by 0.009 AUC.The reported differences are statistically significant (p < 0.01 and p < 0.05, respectively); direct comparison with other DDSM work is difficult because training/testing splits are not standardized.Source: §3.3 · pp. 6–7
4 Conclusion
The paper addresses mammogram data scarcity and class imbalance with a multi-scale, class-conditional GAN that infills masks. It reports realistic lesion generation and improved classification performance over traditional augmentation.
4 Conclusion
- The conclusion frames scarcity of data, especially cancer-positive exams, as a major constraint on progress in mammogram cancer detection.It asks whether GAN-generated lesion examples can synthetically increase training data.Source: §4 · p. 7
- The paper uses a multi-scale class-conditional GAN with mask infilling, called ciGAN, to generate mammogram lesions.The conclusion says ciGAN addresses training instability and resolution detail issues in other GAN architectures.Source: §4 · p. 7
- The conclusion reports that generated realistic lesions improve subsequent classification performance above traditional augmentation techniques.It suggests the techniques may help address data scarcity and class imbalance in a variety of medical imaging settings.Source: §4 · p. 7