Abstract
Serial chest X-rays (CXRs) are routinely compared to monitor disease progression and treatment response, yet the same ordered prior-current pair may require different analyses depending on the user question. Existing longitudinal CXR models generally follow a fixed prediction pathway, while general CXR agents can invoke multiple tools but do not maintain ordered same-patient pair comparison as their primary reasoning context. We propose LongCXR-Agent, a longitudinal CXR agent that maps each free-text query to one of four comparison types and selectively invokes five specialized tools for presence transition, severity progression, spatial evidence, and structural comparison. On a 666-episode longitudinal question-answering benchmark, LongCXR-Agent outperforms direct 4B multimodal language models, including MedGemma, yielding a 0.145 absolute improvement in exact-match accuracy over the best-performing baseline. Query-conditioned spatial evidence further improves progression macro-F1 from 59.4% to 64.0% while reducing latency-weighted execution cost compared with invoking the localization tool for every case. Our experiments demonstrate that LongCXR-Agent preserves temporal order, separates distinct comparison tasks, and acquires additional evidence according to the user query.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_024.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=33vEcuxdrP
BibTex
@InProceedings{JunYoo_LongCXRAgent_MICCAISAT2026,
author = { Jung, Yoonsung AND Kim, Yeonghyeon AND Park, Chang Min AND Lee, Dongheon},
title = { { LongCXR-Agent: Query-Driven Tool Selection for Longitudinal Chest X-ray Reasoning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}