Medicine

Self-supervision drives representational convergence in medical foundation models more than clinical supervision

Soroosh Tayebi Arasteh, Sebastian Ziegelmayer, Mahshad Lotfinia, Lisa Adams, Sven Nebelung, Jakob Nikolas Kather, Daniel Truhn

Featured August 3, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Medical AI models learn to 'see' diseases similarly mostly through self-supervised training, not just from being taught with specific labels, which helps them share tools across different hospitals.

In depth

This paper rigorously demonstrates that self-supervised objectives, rather than clinical supervision or model scale, are the primary drivers of representational convergence in medical foundation models. Despite this convergence being modest and within-modality, the authors show it is sufficient for practical functional interchangeability, enabling diagnostic classifiers to transfer across different encoders and clinical sites.

Key Takeaways

1
The self-supervised objective is the dominant factor driving representational convergence in medical image encoders, not clinical supervision or model size.
2
Representational convergence is a within-modality phenomenon that does not extend to clinical language and is uneven across patient subgroups.
3
Despite modest geometric alignment, the shared representation enables functional interchangeability, allowing diagnostic classifiers and features to transfer effectively across encoders and sites.

Conceptual Flow

HIGH LEVEL

Methodology: How models were compared

The study checked if different medical AI models, trained in various ways, organized information about images in similar internal patterns.

Medical Images

Extract Features

Model A's View

Model B's View

Results: What drives similarity

They found that models learning on their own (self-supervised) became much more similar than those taught with specific disease names.

Self-Supervised Training

Label-Supervised Training

Leads To

High Similarity

Low Similarity