Learning to Unlearn: Building Immunity to Dataset Bias in Medical Imaging Studies

Ahmed Ashraf, Shehroz S. Khan, Nikhil Bhagwat, M. Mallar Chakravarty, Babak TaatiarXiv 2018 · OpenReview

TL;DR

This paper tackles the critical issue of "dataset bias" in medical imaging AI, where models trained on one dataset fail to generalize to others due to dataset-specific quirks. They propose a novel "bias unlearning" framework that creates an intermediate representation space where data from different studies becomes indistinguishable, thereby forcing models to learn fundamental disease characteristics rather than dataset artifacts.

What Problem Does This Paper Address?

Medical imaging AI models are typically trained and tested on data from a single study. While this might seem sufficient, models often struggle to generalize to data from different studies, even when those studies investigate the same medical condition. This phenomenon, known as "dataset bias," means that models might be learning superficial, dataset-specific patterns (like variations in scanner settings or patient demographics unique to a study) rather than the true underlying biological signals. This severely limits the real-world applicability of these AI tools, as they can't reliably be transferred between different clinical settings or research cohorts. The authors highlight that this problem, well-recognized in general computer vision, has been largely overlooked in medical imaging research.

Methods

The core of the paper is a new framework designed to "unlearn" dataset membership. They propose mapping data from its original feature space into an intermediate "latent space." The key idea is to learn "prototypes" in this latent space such that data points mapped to the same prototype become indistinguishable in terms of which original dataset they came from. This is achieved by maximizing the entropy of the distribution of dataset memberships given a prototype assignment, effectively making the dataset origin information "uniform" and thus uninformative.

To prevent trivial solutions (e.g., mapping all data to a single, uninformative prototype), they introduce two additional constraints:

  1. Reconstruction Constraint: Prototypes are learned such that the original data can be approximately reconstructed from their latent space representation. This encourages prototypes to retain meaningful information.
  2. Classification Loss: A standard cross-entropy loss is used to ensure that the model can still accurately classify the medical condition of interest (e.g., distinguishing between healthy individuals and those with Mild Cognitive Impairment).

These three objectives (bias unlearning, reconstruction, and classification) are combined into a single loss function that is optimized to learn the prototypes and classification parameters.

Datasets

The authors evaluated their method on four brain MRI datasets related to cognitive health:

  • The Alzheimer's Disease Neuroimaging Initiative (ADNI) datasets (ADNI 1, ADNI 2, ADNI GO).
  • The Australian Imaging, Biomarker and Lifestyle Flagship Study of Aging (AIBL).

Crucially, for the ADNI studies, they ensured no participant overlap between the datasets. The imaging features used were derived from 78 regions of interest (ROIs) based on the Automatic Anatomical Labeling (AAL) atlas.

Key Insights and Ideas

  • Dataset bias is a real and significant problem in medical imaging: The paper provides empirical evidence that models trained on one medical imaging dataset struggle to generalize to others, even when the underlying task is the same.
  • "Unlearning" dataset membership is key to generalization: By explicitly forcing models to ignore dataset origin, they can learn more robust and generalizable representations.
  • Fairness and generalization are linked: The proposed method draws inspiration from techniques used in privacy-preserving machine learning (learning fair representations), suggesting a broader connection between these fields.
  • Prototypes as a bridge: The learned prototypes act as an intermediate representation that disentangles dataset-specific information from task-relevant information.

Key Results

  • Improved Cross-Dataset Generalization: The proposed "bias unlearning" method significantly improved cross-dataset generalization performance (measured by AUC) compared to naive leave-one-dataset-out (LODO) strategies and an "unbiased SVM" baseline. On average, AUC improved by 47 and 18 points, respectively.
  • Reduced Performance Drop: When comparing cross-dataset performance to within-dataset performance (the "ceiling"), the proposed method showed a much smaller drop in AUC (less than 20%) compared to baselines.
  • "Name the Study" Classifier: An initial experiment showed that a simple classifier could identify the dataset of origin for MRI scans with 65.92% accuracy, confirming that dataset-specific artifacts are present.

Limitations and Future Work

  • Hyperparameter Tuning: The method involves four hyperparameters that require tuning, which was performed using a nested LODO strategy. The authors suggest these could be determined via cross-validation.
  • ADNI-GO Specificity: The ADNI-GO dataset presented some challenges as it contained examples from only one class, impacting AUC computation in certain LODO scenarios.
  • Future Work: The paper implicitly suggests that this framework can be extended to other medical imaging modalities and tasks. The connection to fair machine learning also opens avenues for further research.

Summary updated

Abstract

Medical imaging machine learning algorithms are usually evaluated on a single dataset. Although training and testing are performed on different subsets of the dataset, models built on one study show limited capability to generalize to other studies. While database bias has been recognized as a serious problem in the computer vision community, it has remained largely unnoticed in medical imaging research. Transfer learning thus remains confined to the re-use of feature representations requiring re-training on the new dataset. As a result, machine learning models do not generalize even when trained on imaging datasets that were captured to study the same variable of interest. The ability to transfer knowledge gleaned from one study to another, without the need for re-training, if possible, would provide reassurance that the models are learning knowledge fundamental to the problem under study instead of latching onto the idiosyncracies of a dataset. In this paper, we situate the problem of dataset bias in the context of medical imaging studies. We show empirical evidence that such a problem exists in medical datasets. We then present a framework to unlearn study membership as a means to handle the problem of database bias. Our main idea is to take the data from the original feature space to an intermediate space where the data points are indistinguishable in terms of which study they come from, while maintaining the recognition capability with respect to the variable of interest. This will promote models which learn the more general properties of the etiology under study instead of aligning to dataset-specific peculiarities. Essentially, our proposed model learns to unlearn the dataset bias.
BibTeX
@misc{ashraf2018learning,
  author = {Ahmed Ashraf and Shehroz S. Khan and Nikhil Bhagwat and M. Mallar Chakravarty and Babak Taati},
  title = {{Learning to Unlearn: Building Immunity to Dataset Bias in Medical Imaging Studies}},
  year = {2018},
  eprint = {1812.01716},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.01716},
}
Figure