[Paper Review] Adjusting for Confounding in Unsupervised Latent Representations of Images

Craig A. Glastonbury, Michael Ferlaino+2 more|arXiv (Cornell University)|Nov 15, 2018

Cell Image Analysis Techniques10 references4 citations

TL;DR

This paper proposes an adversarial training method to learn unsupervised, confounder-invariant representations from biomedical images by disentangling nuisance factors like batch effects and staining variability. Using a deep convolutional autoencoder with an adversarial discriminator, the model achieves latent representations that are invariant to confounders while preserving biological information, as shown by 55.4% MOA prediction accuracy at λ=50, significantly above random chance.

ABSTRACT

Biological imaging data are often partially confounded or contain unwanted variability. Examples of such phenomena include variable lighting across microscopy image captures, stain intensity variation in histological slides, and batch effects for high throughput drug screening assays. Therefore, to develop "fair" models which generalise well to unseen examples, it is crucial to learn data representations that are insensitive to nuisance factors of variation. In this paper, we present a strategy based on adversarial training, capable of learning unsupervised representations invariant to confounders. As an empirical validation of our method, we use deep convolutional autoencoders to learn unbiased cellular representations from microscopy imaging.

Motivation & Objective

• To develop unsupervised representation learning methods that are invariant to nuisance factors in biomedical image data, such as batch effects and staining variability.
• To address the problem of poor generalization in machine learning models due to confounding factors in image data, especially in high-throughput screening and histology.
• To enable fair and unbiased representation learning without requiring expensive human annotations, supporting discovery of novel cellular phenotypes.
• To generalize adversarial debiasing from supervised to unsupervised representation learning in image data.

Proposed method

• A deep convolutional autoencoder (CAE) is trained to reconstruct input images from a learned latent code, capturing meaningful biological features.
• An adversarial discriminator is trained to predict the confounder (e.g., batch or staining intensity) from the latent code, encouraging the CAE to produce confounder-invariant representations.
• The joint training objective combines CAE reconstruction loss and adversarial loss with a hyperparameter λ balancing the two, minimizing the CAE loss while maximizing the adversarial loss.
• The optimization is framed as a min-max problem: minimize the CAE parameters θ while maximizing the discriminator weights w, ensuring the latent code is uninformative about the confounder.
• The method is extended to continuous confounders by using an adversarial regressor instead of a classifier.
• The approach is validated on the BBBC021v1 microscopy dataset, where it successfully reduces batch effect influence while preserving biological signal for MOA prediction.

Experimental results

Research questions

• RQ1 Can adversarial training be adapted to learn unsupervised, confounder-invariant representations in biomedical image data without requiring labeled annotations?
• RQ2 To what extent can confounding factors like batch effects and staining variability be removed from learned image representations while preserving biologically relevant information?
• RQ3 How does the trade-off between confounder invariance and biological information retention manifest across different values of the adversarial weighting hyperparameter λ?
• RQ4 Can this method generalize to continuous confounders such as ischemic time in histological data?
• RQ5 Does the resulting representation improve downstream classification performance on biologically meaningful tasks like drug mechanism of action prediction?

Key findings

• At λ=50, the model achieves 55.435% accuracy in predicting molecular mechanism of action (MOA), which is 3.6× better than random chance, demonstrating preservation of biologically relevant information.
• The batch prediction accuracy drops to 35.870% at λ=50, indicating strong invariance to the batch effect confounder, which is significantly below the 64.130% accuracy at λ=0.
• The method successfully removes confounding signals from the latent space, as evidenced by the sharp decline in batch accuracy with increasing λ, while maintaining high MOA prediction performance.
• The approach generalizes to continuous confounders: a correlation of r=0.18 (p=4.9×10−150) was found between ischemic time and unsupervised representations in GTEx histology data, confirming the presence of a continuous confounder.
• The model outperforms random chance by over 3× in MOA prediction even at high λ values, confirming that the learned representations retain meaningful biological structure.
• The method enables discovery of novel phenotypes by avoiding reliance on pre-labeled annotations, as it operates in an unsupervised paradigm.