Abstract
Medical visionlanguage models (VLMs) such as MedGemma are trained predominantly on adult data, yet pediatric imaging is underserved precisely where trustworthy automation is most needed. We present a ne-tuning-free, reproducible evaluation of MedGemma on the pediatric subset (patients aged < 18, n=1,276; 385 pneumonia) of the RSNA (i) Zero-shot classi cation: MedGemma reaches 79.4% accuracy for binary pneumonia detecPneumonia Detection dataset, along three axes. tion but only 59.2% three-class (κ=0.38), a non-trivial yet limited baseline. (ii) Age-matched retrieval: training-free retrieval-augmented generation with a pediatric image index yields the only statistically significant gains in each taskbinary for MedGemma-4B (+5.0 pts, McNemar p=0.004) and three-class for MedGemma-4B-1.5 (+8.3 pts, p=0.006) (iii) Localisation faithfulness: scoring Captum grid attribution and the and never underperforms an adult index, which can itself hurt. model’s predicted box against radiologist opacity masks reveals high presence accuracy but poor spatial overlap (mean boxGT IoU 0.07, below size-matched random (0.16) and centered (0.19) boxes), i.e. right-answer / wrong-region behaviour. All comparisons carry bootstrap 95% CIs and hold across two MedGemma versions. We conclude that age-matched retrieval is a cheap, eective pediatric adaptation, whereas spatial unfaithfulness is a concrete safety caution for deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/PedAItrics_016.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{ZegRac_Do_MICCAISAT2026,
author = { Zeghlache, Rachid AND Al Musleh, Mohamed AND Ahmed, Rehan AND Oroumchian, Farhad AND Alshryda, Sattar AND Elserafi, Ahmed AND Abdulaziz, Nidhal},
title = { { Do Medical Vision-Language Models Work for Children? Evaluating MedGemma on Pediatric Pneumonia } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17257},
month = {pending},
page = {pending}
}