Source-linked AI summary
Synthetic Data Augmentation using GAN for Improved Liver Lesion Classification
Maayan Frid-Adar, Eyal Klang, Michal Amitai, Jacob Goldberger, Hayit Greenspan
TL;DR
Limited medical-imaging datasets and expert annotations motivate augmenting liver-lesion classification data beyond classical transformations. The paper combines classical augmentation with GAN-generated synthetic lesions, improving classification performance over classical augmentation alone.
Problem
Medical imaging classification is constrained by small datasets and limited annotated samples, while liver CT images require automated diagnostic support.
Method
The method first applies classical augmentation, then trains class-specific GANs to generate synthetic lesions for additional CNN training augmentation.
Results
Synthetic augmentation improved classification accuracy from 78.6% with classical augmentation alone to 85.7%.
Takeaways & Limitations
The study demonstrated a 7% improvement using synthetic augmentation over classic augmentation for liver lesion classification.
Abstract
from arXiv · showhide
In this paper, we present a data augmentation method that generates synthetic medical images using Generative Adversarial Networks (GANs). We propose a training scheme that first uses classical data augmentation to enlarge the training set and then further enlarges the data size and its diversity by applying GAN techniques for synthetic data augmentation. Our method is demonstrated on a limited dataset of computed tomography (CT) images of 182 liver lesions (53 cysts, 64 metastases and 65 hemangiomas). The classification performance using only classic data augmentation yielded 78.6% sensitivity and 88.4% specificity. By adding the synthetic data augmentation the results significantly increased to 85.7% sensitivity and 92.4% specificity.
1. INTRODUCTION
Medical imaging classifiers face small datasets and limited expert annotations, motivating augmentation methods that add variability beyond basic image modifications. This paper applies GAN-based synthetic lesion generation to liver lesion classification.
- Motivation: Small medical-imaging datasets and limited annotated samples challenge supervised learning.Annotations require radiologist expertise and are time consuming, while collecting medical data is complex and expensive.
- Motivation: Classical augmentation uses image modifications such as translation, rotation, flipping, and scaling, but small modifications provide relatively limited diversity.
- Motivation: Synthetic examples can introduce more variability and further enrich limited datasets for model training.
- Approach: The study investigates GANs for synthesizing high-quality medical images to augment data for liver lesion classification.The work focuses on CT-based classification of focal liver lesions.
- Clinical motivation: Automated CT-based liver lesion classification addresses the need for diagnostic tools that assist radiologists.The introduction identifies the liver as one of the three most common sites for metastatic cancer.
- Contribution: The proposed scheme combines standard image perturbation, GAN-generated liver lesions, and CNN classification to improve classification results.
2. GENERATING SYNTHETIC LIVER LESIONS
The method first enlarges lesion data with classical transformations, then trains GANs to generate additional class-specific synthetic lesions for classifier training. GANs learn image distributions through adversarial generator–discriminator training.
- Proposed pipeline: The pipeline first applies standard augmentation, trains a GAN on the enlarged dataset, and uses generated examples to augment lesion-classifier training.
- Classical augmentation: Classical augmentation uses rotations and flips while preserving lesion characteristics by avoiding shape-deforming transformations and keeping the ROI centered.
- GAN framework: GANs implicitly learn the data distribution from samples and generate new samples drawn from the learned distribution.
- GAN framework: A DCGAN synthesizes labeled lesions separately for cysts, metastases, and hemangiomas using two simultaneously trained deep CNNs.
- Examples: Figure 1 contrasts real and synthetic lesion ROIs across cyst, metastasis, and hemangioma classes.
- GAN framework: The generator maps random input samples to image space while the discriminator estimates whether an input is real, enabling adversarial training.
- Network implementation: The generator outputs 64 × 64 × 1 liver lesion images from vectors of 100 uniformly distributed random numbers.
3. EXPERIMENTS AND RESULTS
The experiments evaluate classical and GAN-based augmentation for three-class liver lesion classification on 182 annotated CT lesions. Synthetic augmentation improved classification beyond the classical-augmentation baseline, while radiologists rated real and synthesized lesions similarly.
- Data and Implementation: The dataset contained 182 portal-phase 2-D CT scans: 53 cysts, 64 metastases, and 65 hemangiomas, with biopsy or clinical follow-up ground truth.
- Data and Implementation: The classifier used 64×64 lesion ROIs, a CNN with convolutional, pooling, dense, dropout, and soft-max layers, evaluated using patient-level 3-fold cross-validation.
- Classical Data Augmentation: The CNN-AUG baseline enlarged each original lesion ROI through rotation, flipping, translation, and scaling, producing 480 augmented images per lesion ROI.
- Synthetic Data Augmentation: The CNN-AUG-GAN procedure trained separate DCGANs for cysts, metastases, and hemangiomas, using class-balanced synthetic ROIs while maintaining train-test separation.
- Classification Results: Classification accuracy increased from 78.6% with classical augmentation to 85.7% with synthetic lesions after classical augmentation saturated.
- Expert Assessment: Two radiologists classified real and synthesized lesions similarly, with Expert 1 achieving 78% versus 77.5% and Expert 2 achieving 69.2% versus 69.2%.
4. CONCLUSION
The work uses synthetic medical-image generation for data augmentation in liver-lesion classification with limited data, improving performance over classic augmentation.
- The method generates synthetic medical images to augment training for liver-lesion classification with limited data.
- 7% improvement was achieved using synthetic augmentation over classic augmentation.
- The authors plan to extend the approach to additional medical domains that may benefit from lesion synthesis for improved training.