Exploring Prompt Alignment with Clinical Factors in Zero-Shot Segmentation VLMs for NSCLC Tumor Segmentation
Suraj Pai, Thibault Heintz, Cosmin Ciausu, Marion Tonneau, Hugo Aerts, Raymond Mak · 2026
Read Exploring Prompt Alignment with Clinical Factors in Zero-Shot Segmentation VLMs for NSCLC Tumor Segmentation by Suraj Pai, Thibault Heintz, Cosmin Ciausu, Marion Tonneau, Hugo Aerts, Raymond Mak (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-01266.txt.
Description
Suraj Pai and five co-authors explore how prompt wording aligns with clinical factors in zero-shot segmentation models for lung cancer tumors. Promptable vision-language models offer an alternative to task-specific training for delineating gross tumor volume in non-small-cell lung cancer, but which prompt dimensions govern spatial behavior is poorly understood. The team splits prompts into diagnosis, demographic, staging, anatomical, generic, and irrelevant parts on a held-out dataset.
Anatomical location emerges as the dominant driver of spatial attention. Most location perturbations cause catastrophic performance drops, while histology and stage substitutions matter little. Irrelevant prompts correctly yield nothing. Swapped prompts confirm per-patient conditioning. The zero-shot model studied is statistically indistinguishable from strong fine-tuned baselines on overlap score, while beating other zero-shot models.
Catalog record
| CREATOR | Suraj Pai, Thibault Heintz, Cosmin Ciausu, Marion Tonneau, Hugo Aerts, Raymond Mak |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 16 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.01266