Lessons from shortcomings in machine learning for medical imaging

Gaël Varoquaux, Veronika CheplyginaOECD 2023 · DOI · OpenAlex

Abstract: The application of machine learning (ML) to medical imaging has attracted a lot of attention in recent years. Yet, for various reasons, progress remains slow. This essay builds upon earlier work by the authors which explores how larger datasets and more deep-learning algorithms have not yet provided practical improvements in addressing clinical problems. It recommends how researchers and policy makers can improve the situation.

Introduction: The Clinical Impact Gap

Despite the proliferation of "expert-level" AI claims in medical literature, a significant "clinical impact gap" persists. While thousands of papers report high diagnostic accuracy, few tools achieve clinical deployment. For example, a review of 62 machine learning studies for COVID-19 found that none were fit for clinical use. This discrepancy suggests that current research incentives prioritize flashy "state-of-the-art" (SOTA) results over the rigorous validation required for real-world medicine.

Methodology

This perspective builds on a meta-analysis and empirical critique of the field. The authors conducted two primary investigations:

  1. Alzheimer’s Research Meta-analysis: A review of over 500 publications to track reported accuracy against study sample sizes.
  2. Kaggle Competition Analysis: An investigation of four major medical imaging competitions to compare the performance gains of top models against "evaluation noise" (the variability between public and private leaderboards).

The Large Dataset Paradox

A critical finding is that studies with larger sample sizes tend to report lower accuracy. Smaller studies often suffer from "optimism bias" or overfitting to small, homogeneous, and clean cohorts. In contrast, larger datasets introduce the messiness of real-world clinical conditions, revealing that the diagnostic tasks are actually much harder than small-scale papers suggest.

Reported accuracy vs. study size in Alzheimer's research shows a downward trend as cohorts become larger and more realistic.

As shown in the analysis of Alzheimer's research, accuracy for clinically relevant tasks—such as distinguishing progressive from stable Mild Cognitive Impairment (pMCI vs. sMCI)—has remained stagnant for years despite the availability of more data and more complex models.

Evaluation Noise and Competition

The obsession with leaderboard rankings often masks statistical insignificance. In three out of four analyzed Kaggle competitions, the margin by which the "winning" model outperformed the top 10% of participants was smaller than the random noise inherent in the evaluation set.

In many competitions, the improvement of the top model is indistinguishable from evaluation noise.

This suggests that "winning" models may simply be those that best fit the specific noise of a test split, rather than those offering superior clinical generalization. We are effectively rewarding researchers for winning a coin flip.

Systemic Barriers: Compute and Incentives

The field faces a growing "compute divide." At MICCAI 2020, only 2% of papers originated from underrepresented regions like Africa or Latin America, where the need for automated medical assistance is often highest. Furthermore, the pressure to publish "novel" algorithms leads to over-engineered models that solve specific benchmarks but fail to address broader clinical problems.

Recommendations for a More Reliable Future

To bridge the impact gap, the authors propose several shifts in research culture:

  • Reinventing Benchmarking: Move beyond single accuracy metrics toward reporting performance ranges across diverse patient groups and real-world utility.
  • Registered Reports: Adopt peer review for study designs before experiments are run to reduce the pressure to "massage" data for high accuracy scores.
  • Policy and Funding: Prioritize the curation of high-quality, open datasets and the rigorous validation of existing models over the constant pursuit of algorithmic novelty.

By shifting the goalposts from novelty to reliability, the medical AI community can ensure that its significant investments lead to tangible improvements in patient care rather than academic noise.

Summary updated

Abstract

The application of machine learning (ML) to medical imaging has attracted a lot of attention in recent years. Yet, for various reasons, progress remains slow. This essay builds upon earlier work by the authors which explores how larger datasets and more deep-learning algorithms have not yet provided practical improvements in addressing clinical problems. It recommends how researchers and policy makers can improve the situation.
BibTeX
@inproceedings{varoquaux2023lessons,
  author = {Gaël Varoquaux and Veronika Cheplygina},
  title = {{Lessons from shortcomings in machine learning for medical imaging}},
  booktitle = {OECD},
  year = {2023},
  doi = {10.1787/b885eecd-en},
}

Loading related content...

Figure