Conditional Infilling GANs for Data Augmentation in Mammogram Classification

Eric Wu, Kevin Wu, David D. Cox, William Lotter

TL;DR

This paper introduces a novel Conditional Infilling GAN (ciGAN) to synthetically generate realistic mammogram patches with synthesized lesions. By augmenting existing datasets with these GAN-generated images, they demonstrate a significant improvement in breast cancer classification performance compared to traditional augmentation methods.

What Problem Does This Paper Address?

Deep learning models have shown great promise in detecting breast cancer from mammograms. However, their performance is often hampered by the limited availability of large, annotated datasets. This scarcity is due to privacy concerns, the high cost of expert annotations, and a significant class imbalance where normal mammograms vastly outnumber those with abnormalities. Traditional data augmentation techniques, like flipping and rotation, help but may not be sufficient to overcome these limitations. The authors explore whether Generative Adversarial Networks (GANs) can be leveraged to create more sophisticated synthetic data to address this challenge.

Methods

The core of the paper is the Conditional Infilling GAN (ciGAN). This architecture is designed to synthesize realistic mammogram patches by "infilling" a specific region. Key aspects of their approach include:

  • Multi-scale Generator: The generator uses a cascading refinement network that generates features at multiple scales, improving stability for high-resolution outputs.
  • Contextual Infilling: Instead of generating an entire image, the generator is constrained to fill in only a segmented lesion area, making the task more manageable and focused.
  • Class-Conditional: The GAN is conditioned on class labels (malignant/non-malignant), allowing it to generate specific types of lesions or normal tissue.
  • Input Structure: The input to the generator is a stack of channels: a corrupted image (where the lesion area is replaced with noise), a mask indicating the lesion's location, and class labels.
  • Loss Functions: The training employs a combination of:
  • Feature Loss: Uses a pre-trained VGG-19 network to match feature maps between real and generated images, ensuring visual realism.
  • Adversarial Loss: The standard GAN loss to ensure generated images are indistinguishable from real ones.
  • Boundary Loss: Encourages smooth transitions between the synthesized lesion and the surrounding healthy tissue.

Datasets

  • DDSM Dataset: The primary dataset used is the Digital Database for Screening Mammography (DDSM). It contains 10,480 total images, with 17.5% malignant and 82.5% non-malignant cases. The authors split this dataset into 80% training, 10% validation, and 10% testing. They also generated 100,000 random 256x256 patches, accepting only those with over 75% breast tissue.

Key Insights and Ideas

  • GANs for Targeted Augmentation: The paper demonstrates that GANs can be effectively used not just for general image generation but for targeted data augmentation by synthesizing specific abnormalities (lesions) onto healthy images.
  • Contextual Infilling is Key: Constraining the GAN to infill only the lesion area, rather than the whole image, simplifies the generation process and allows for more realistic synthesis of localized findings.
  • Improved Classification with Synthetic Data: The results strongly suggest that GAN-generated data, when combined with traditional augmentation, leads to a tangible improvement in the performance of a ResNet-50 classifier for mammogram classification.
  • Addressing Data Scarcity and Imbalance: The proposed method offers a promising avenue to mitigate the critical issues of limited data and class imbalance in medical imaging.

Key Results

  • Qualitative: The ciGAN successfully generates high-resolution (256x256 pixels) synthetic mammogram patches, with generated lesions that appear realistic and well-integrated into the surrounding tissue.
  • Quantitative:
    • A ResNet-50 classifier trained with ciGAN + traditional augmentation achieved an AUC of 0.896.
  • This is a significant improvement over a baseline model with no augmentation (AUC of 0.882) and a model with only traditional augmentation (AUC of 0.887).
  • The ciGAN-augmented model outperformed the traditional augmentation model by 0.009 AUC (p < 0.05).

Limitations and Future Work

  • Limitations:
    • The paper notes that direct comparison with other works is challenging due to the lack of standardized training/testing splits for the DDSM dataset.
  • While the results are promising, the authors acknowledge that their work is a "first step" towards harnessing GAN-based data augmentation for mammography.
  • Future Work:
    • The authors believe their techniques can be applied to address data scarcity and class imbalance in a variety of medical imaging settings.
  • Further exploration of GAN architectures and training strategies for medical imaging is implied.

Summary updated

BibTeX
@inproceedings{wu2018conditional,
  author = {Eric Wu and Kevin Wu and David D. Cox and William Lotter},
  title = {{Conditional Infilling GANs for Data Augmentation in Mammogram Classification}},
  booktitle = {RAMBO+BIA+TIA @ MICCAI},
  year = {2018},
  url = {https://openreview.net/forum?id=St8bOPGoEA},
}

Loading related content...

Figure