Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TSynD: Targeted Synthetic Data Generation for Enhanced Medical Image Classification

Joshua Niemeijer, Jan Ehrhardt, Hristina Uzunova, Heinz Handels

TL;DR

Medical imaging datasets are scarcity-challenged and annotation-intensive. TSynD addresses this by using latent-space optimization of an autoencoder to generate targeted synthetic samples that maximize epistemic uncertainty, thereby covering missing regions of the data distribution. Empirical results on MedMNIST v2 and Chest-Xray show that TSynD improves accuracy and robustness under low-data conditions and test-time perturbations, outperforming baseline and random latent-space augmentation, with MI-based uncertainty yielding stronger gains than entropy. Qualitative analyses indicate more meaningful regions are utilized and that generated samples contribute to greater resilience against adversarial perturbations.

Abstract

The usage of medical image data for the training of large-scale machine learning approaches is particularly challenging due to its scarce availability and the costly generation of data annotations, typically requiring the engagement of medical professionals. The rapid development of generative models allows towards tackling this problem by leveraging large amounts of realistic synthetically generated data for the training process. However, randomly choosing synthetic samples, might not be an optimal strategy. In this work, we investigate the targeted generation of synthetic training data, in order to improve the accuracy and robustness of image classification. Therefore, our approach aims to guide the generative model to synthesize data with high epistemic uncertainty, since large measures of epistemic uncertainty indicate underrepresented data points in the training set. During the image generation we feed images reconstructed by an auto encoder into the classifier and compute the mutual information over the class-probability distribution as a measure for uncertainty.We alter the feature space of the autoencoder through an optimization process with the objective of maximizing the classifier uncertainty on the decoded image. By training on such data we improve the performance and robustness against test time data augmentations and adversarial attacks on several classifications tasks.

TSynD: Targeted Synthetic Data Generation for Enhanced Medical Image Classification

TL;DR

Medical imaging datasets are scarcity-challenged and annotation-intensive. TSynD addresses this by using latent-space optimization of an autoencoder to generate targeted synthetic samples that maximize epistemic uncertainty, thereby covering missing regions of the data distribution. Empirical results on MedMNIST v2 and Chest-Xray show that TSynD improves accuracy and robustness under low-data conditions and test-time perturbations, outperforming baseline and random latent-space augmentation, with MI-based uncertainty yielding stronger gains than entropy. Qualitative analyses indicate more meaningful regions are utilized and that generated samples contribute to greater resilience against adversarial perturbations.

Abstract

The usage of medical image data for the training of large-scale machine learning approaches is particularly challenging due to its scarce availability and the costly generation of data annotations, typically requiring the engagement of medical professionals. The rapid development of generative models allows towards tackling this problem by leveraging large amounts of realistic synthetically generated data for the training process. However, randomly choosing synthetic samples, might not be an optimal strategy. In this work, we investigate the targeted generation of synthetic training data, in order to improve the accuracy and robustness of image classification. Therefore, our approach aims to guide the generative model to synthesize data with high epistemic uncertainty, since large measures of epistemic uncertainty indicate underrepresented data points in the training set. During the image generation we feed images reconstructed by an auto encoder into the classifier and compute the mutual information over the class-probability distribution as a measure for uncertainty.We alter the feature space of the autoencoder through an optimization process with the objective of maximizing the classifier uncertainty on the decoded image. By training on such data we improve the performance and robustness against test time data augmentations and adversarial attacks on several classifications tasks.

Paper Structure

This paper contains 13 sections, 5 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Estimation of the epistemic uncertainty
  5. Targeted Synthetic Data Generation
  6. Latent space noise.
  7. The training process.
  8. Optimizing latent codes vs. pixel values as parameters.
  9. Experiments
  10. Classification and robustness results
  11. Uncertainty Maximization
  12. Qualitative Robustness Evaluation
  13. Conclusion

Figures (2)

  • Figure 1: The overall framework of TSynD (Targeted Synthetic Data generation) for the robust training of a classifier: During classifier training, the latent spatial representations of original images are optimized to maximize the classifier's epistemic uncertainty in the decoded images. The new images then serve as additional training data.
  • Figure 2: Left: EigenGradCAM maps of the baseline classifier and classifier trained with TSynD. Right: Perturbation of images to minimize the probability of the given class. Depicted is the difference of the images at the start and end of the minimization.