TLDR: GeMix is a novel data augmentation framework that uses conditional GANs to create realistic and semantically consistent synthetic medical images. Unlike traditional mixup which blends pixels, GeMix interpolates class labels to condition a generator, producing anatomically plausible samples. Benchmarked on the COVIDx-CT-3 dataset, GeMix significantly improves classification performance and reduces false negatives for COVID-19 detection when combined with real data, outperforming conventional mixup methods.
Deep learning models have achieved remarkable success in image classification, but they often require vast amounts of labeled data and robust training techniques. Data augmentation, a strategy that generates new training examples from existing ones, helps address these challenges. One popular augmentation method is ‘mixup’, which creates synthetic images by linearly interpolating pairs of training examples and their labels. While effective for improving generalization, traditional mixup often produces visually unrealistic images, especially in sensitive fields like medical imaging where subtle details carry significant diagnostic meaning. These unrealistic images can sometimes hinder the learning process.
Introducing GeMix: A Smarter Approach to Augmentation
To overcome the limitations of traditional mixup, researchers have introduced GeMix, a novel two-stage framework designed for improved medical image augmentation. GeMix replaces the simple pixel-wise blending of mixup with a learned, label-aware interpolation process powered by class-conditional Generative Adversarial Networks (GANs). This allows GeMix to synthesize visually coherent and anatomically plausible images that lie along a continuous class manifold, rather than creating ambiguous, blended artifacts.
How GeMix Works
The GeMix pipeline operates in two main stages. First, a StyleGAN2-ADA generator, a type of advanced GAN, is trained on the specific medical image dataset. This generator learns to produce realistic images conditioned on class information. In the second stage, during the augmentation process, instead of directly mixing images, GeMix samples two label vectors from Dirichlet priors. These priors are biased towards different classes, and they are blended using a Beta-distributed coefficient. The generator then uses this ‘soft label’ (a mixed class label) to synthesize new images. Because the interpolation happens in the label space that conditions the GAN, the resulting images are much more realistic and semantically meaningful than those produced by pixel-level blending.
Benchmarking GeMix on Medical Data
The effectiveness of GeMix was rigorously evaluated on the large-scale COVIDx-CT-3 dataset, which comprises over 430,000 CT slices from thousands of patients, covering COVID-19, community-acquired pneumonia (CAP), and normal cases. The researchers tested GeMix with three widely used deep learning architectures: ResNet-50, ResNet-101, and EfficientNet-B0. They compared GeMix against traditional mixup and a generalized multi-image mixup (MMixup) under various training scenarios, including using only augmented data and combining augmented data with real images.
Also Read:
- Unlocking AI’s Memory: How Synthetic Images Combat Forgetting in Learning Systems
- Smart Labeling: How ConformalSAM Improves Segmentation with Foundational Models
Key Findings and Benefits
The results demonstrated that when combined with real data, GeMix consistently improved macro-F1 scores across all tested backbones compared to traditional mixup and MMixup. For instance, with ResNet-101, the Real+GeMix setup achieved the best performance, significantly surpassing other configurations. A crucial benefit observed was a reduction in the false negative rate for COVID-19 detection, meaning the model was better at identifying positive cases. This is particularly important in high-stakes medical applications where missing a diagnosis can have severe consequences.
Qualitative analysis further highlighted GeMix’s advantages. Unlike traditional mixup and MMixup, which often produce images that are not anatomically plausible, GeMix generated images that were visually realistic and coherent. This is because GeMix performs interpolation in the class-conditioning input of the GAN, rather than directly on the image pixels. Furthermore, GeMix was shown to expand the data distribution more effectively than other augmentation methods, leading to a richer and more diverse training set for the models.
In conclusion, GeMix offers a principled and effective alternative to pixel-space mixup for medical image augmentation. By leveraging conditional GANs and soft-label interpolation, it delivers stronger regularization and greater semantic fidelity without disrupting existing training pipelines. The researchers have made their code publicly available to encourage reproducibility and further research, which can be found at the project’s GitHub repository.
