Incoprea Theory
Volume VI · 2026

Theory 06/Radiology & Medical Imaging /Medical AI/3 min

Medical AI Fails on Physics, Not People

Model drift is blamed on demographics. The images may just look different to the machine.

The claim

Imaging models lose accuracy across sites because of scanner-level differences in acquisition physics rather than because of patient population differences, so generalisation should be fixed by harmonising acquisition and recalibrating physics rather than by collecting more diverse patients.

01The evidence

  • Out-of-distribution generalisation for knee calcium deposition under X-ray manufacturer domain shifts has been studied directly, with performance changing across manufacturers at matched patient populations.MEDI 2026
  • Trustworthy medical multimodal models are assessed for stable, calibrated outputs across clinical scenarios, and the failures reported are often acquisition-related rather than demographic.Trustworthy medical MLLMs

02The tension

Model drift is usually addressed by expanding the training set across populations, on the assumption that the gap is demographic. That is expensive, slow, and it does not fix a gap that is caused by the imaging chain.

03The theory

A medical image is a physical measurement. Reconstruction kernels, detector response, beam quality and post-processing differ between manufacturers and between scanners of the same model, and those differences change texture and edge statistics.

Convolutional and transformer models are sensitive to exactly those statistics. If the shift is physical, then harmonising acquisition or matching the physics in training should transfer performance across sites without new patients.

The prediction is that a model trained on one manufacturer should improve on another when the images are harmonised, and that the residual gap should be smaller than the gap explained by acquisition parameters.

The practical consequence is that procurement should require acquisition metadata, and that regulatory evaluation should test across physics rather than only across populations.

04What this is not

Demographic-diversity accounts of model drift

Those accounts attribute drift to patient mix and prescribe broader data. This attributes it to acquisition physics and prescribes harmonisation, which is cheaper and testable.

05Predictions

  • P1Harmonising acquisition should transfer performance across manufacturers.
  • P2Acquisition parameters should explain more variance than demographics.
  • P3The residual gap after harmonisation should be small.

06What would refute it

Performance gap persisting after physics harmonisation would restore the demographic account.

07What it would change

It would change procurement, validation and how drift is corrected.

08Evidence base

  1. Out-of-distribution generalisation under X-ray manufacturer domain shifts — MEDIhttps://dlnext.acm.org/doi/10.1007/978-3-032-19865-5_6
  2. Trustworthy medical multimodal models — calibration across scenarioshttps://zenodo.org/records/22258831/files/Trustworthy_Medical_Multimodal_Large_Language_Models_Arxiv.pdf