Identify the cause of medical tokenization’s departure from the natural-image regime

Identify whether data scarcity, low inter-sample variability, limited texture diversity, or strong anatomical priors is the cause of the observed difference between medical image tokenization and natural-image tokenization.

Background

The study establishes that medical image tokenization differs from assumptions inherited from natural-image research: reconstruction predicts generation, high codebook usage can coexist with collapsed latent geometry, and discretization generally preserves downstream utility. However, the experiments do not establish why these domain-specific behaviors arise.

The authors identify four candidate explanations—data scarcity, low inter-sample variability, limited texture diversity, and strong anatomical priors—and explicitly characterize determining the responsible cause as an open question. Resolving it would provide a mechanistic account of the observed domain shift and guide the design of medical image tokenizers.

References

Finally, our work establishes how medical tokenization departs from the natural-image regime but not yet why; pinning down the cause, be it data scarcity, low inter-sample variability, limited texture diversity, or strong anatomical priors, is the key open question.

— What Makes a Good Medical Image Tokenizer? Rethinking Reconstruction and Generation in Medical Image Tokenization  (2609.24691 - Bubeck et al., 21 Sep 2026) in Section 6, Conclusion, subsection “Limitations and Future Work”