Breast Abnormality Classification in Resource-Constrained Settings Through GAN Generated Synthetic Mammography Patches

Technical Summary: Breast Abnormality Classification in Resource-Constrained Settings Through GAN Generated Synthetic Mammography Patches

Problem Statement
The development of deep learning (DL) models for medical imaging, specifically breast cancer diagnosis via mammography, is frequently hindered by two primary constraints: data scarcity and resource limitations. Breast cancer datasets often exhibit severe class imbalance, with abnormal cases representing approximately 0.5% of screening exams, leading to high rates of false positives and negatives when models are trained on inadequate data. Furthermore, the protection of patient identifiable information (PII) restricts the release of large public datasets, while the computational cost of training generative models often exceeds the capabilities of resource-constrained environments.

Methodology
This study proposes a pipeline to generate synthetic mammography patches using a low-resource Generative Adversarial Network (GAN) to train binary classifiers for abnormality detection. The workflow consists of two phases:

1. Synthetic Image Generation (CUT-GAN):

   • Model: The study utilizes Contrastive Unpaired Translation GAN (CUT-GAN), selected for its low GPU memory consumption and ability to generate high-quality images.
   • Data Source: The Emory Breast Imaging Dataset (EMBED) provided 3.4 million mammography images. Patches were extracted from cohorts 1–8 for training. Positive patches (BI-RADS 0 with radiologist-annotated Regions of Interest) were transformed into synthetic abnormal patches, while negative patches were sourced from non-annotated regions or BI-RADS 1/2 images.
   • Preprocessing: To ensure training stability, patches with low pixel color counts (indicating potential background air or empty patches) were filtered out.
   • Hyperparameter Optimization: The model was trained using the ADAM optimizer with a learning rate of 0.0002, a batch size of 8, and 400 epochs. This configuration was determined to prevent "train failure," which manifested as noise artifacts or empty patches.
   • Post-Processing: Since CUT-GAN generates RGB images while mammography is grayscale, synthetic images were converted to grayscale. Crucially, a window-level adjustment was applied to match the brightness and contrast distributions of the synthetic images to real mammography patches. A Student's t-test confirmed that the adjusted synthetic distributions were statistically indistinguishable from real images (P-value > 0.05).

2. Model Training and Validation:

   • Classifier: Residual Neural Networks (ResNet50, ResNet101, and ResNet152) pre-trained on ImageNet were employed.
   • Architecture: Five additional trainable layers were added to the frozen pre-trained weights. The input layer processed 256x256 pixel patches, utilizing ReLU activation, average pooling, and a sigmoid output for binary classification.
   • Validation: Models trained on the synthetic dataset were validated against real mammography patches from EMBED cohorts 9 and 10 (13,220 total patches). Performance was evaluated using bootstrapping over 500 iterations to estimate 95% confidence intervals.

Key Results
The study compared models trained on synthetic data against those trained on real data:

• Performance Gap: Models trained on real mammography patches demonstrated superior overall performance, achieving an Area Under the Curve (AUC) of approximately 0.91–0.92. In contrast, models trained on synthetic patches achieved a lower AUC, with ResNet152 reaching 0.751 (0.742–0.759).
• Metric Trade-offs: Synthetic-data-trained models exhibited significantly lower accuracy (0.61–0.67) and precision (0.44–0.50) compared to real-data models. However, they demonstrated notably higher recall (0.80–0.81 for ResNet152), suggesting a tendency toward high sensitivity but lower specificity (higher false positive rates).
• Resource Efficiency: The CUT-GAN model was successfully trained using only 3 GB of GPU memory (on a system with 24 GB total capacity), a significant reduction compared to other generative models like diffusion models that often require 24 GB or more.

Significance and Claims
The authors claim that this research demonstrates the feasibility of using low-resource GANs to generate synthetic medical data for training DL models in resource-constrained settings. The study highlights three main contributions:

1. Privacy Preservation: Synthetic patches preserve the structural and textural characteristics of real mammography data without revealing patient-identifiable information, offering a secure alternative for data sharing and model training.
2. Class Imbalance Mitigation: The approach provides a mechanism to generate synthetic samples for underrepresented classes (abnormal cases), addressing the data scarcity inherent in breast cancer screening.
3. Accessibility: By utilizing CUT-GAN, the method enables researchers with limited computational infrastructure to leverage synthetic data augmentation.

The paper concludes that while synthetic-trained models do not yet match the accuracy of those trained on real data, they show potential as substitutes for real patches, particularly in scenarios where privacy or resource constraints prevent the use of large real-world datasets. The authors emphasize that further research is required to optimize generation pipelines and improve the generalizability of these models across different datasets.