# ORGANIC3D Tests Whether CT Models Use the Right Anatomy

> A new framework tests how 3D CT vision-language models link findings to organs, and trains them to rely more on the relevant scan evidence.

Oossa · 2026-11-19 · https://oossa.com/en/organic3d-tests-whether-ct-models-use-the-right-anatomy

Researchers introduced ORGANIC3D, a framework for testing and improving whether AI models that read 3D CT scans connect written findings to the right anatomy. Its tests alter scans by removing organs, replacing lesions or changing left-right orientation, exposing alignment failures even in models with high classification scores.

The researchers say their model, trained exclusively on public data, reached 77% laterality accuracy, versus at most 66% for earlier models. In tests where an organ was removed, it retracted 96–99% of findings about that organ; other models retracted at most 69%.

## The facts

- The benchmark assessed 12 vision encoder checkpoints and 7 generative vision-language models.
- The paper appeared on arXiv on 2026-11-19.

## Why it matters

For radiology teams assessing AI for CT scans, ORGANIC3D offers tests of whether a model’s findings depend on the relevant anatomy, though the paper does not establish how the results translate to clinical use.

## Sources & references

1. [ORGANIC3D: Grounding 3D CT Vision Language Models in Organ Level Image Text Coupling](https://arxiv.org/abs/2611.23529) – arXiv cs.CV, 2026-11-19

Last updated: 2026-11-19
