Abstract
Automated chest X-ray (CXR) report generation requires recognizing spatially localized pathological findings and translating them into radiology language under limited annotated data. Existing methods rely on global visual embeddings, single-view inputs, or decoder-prompted knowledge injection, limiting the structured multi-view reasoning used by radiologists. We present a condition and anatomy-guided multi-view vision-language framework using a unified query-based architecture for limited-data learning. Paired frontal and lateral radiographs are encoded via a shared SigLIP backbone with learned view-type embeddings and adaptive fusion. Fourteen condition-specific query tokens initialized from BioBERT embeddings of CheXbert labels and six anatomical query tokens are jointly optimized through multi-head cross-attention over image patches. An adaptive gated fusion mechanism balances global CLS context with localized query evidence, producing an 11-token visual representation for a Flan-T5 decoder. An auxiliary multi-label classification head provides condition-level supervision, and parameter-efficient LoRA supports domain transfer. Training combines anatomy-aware augmentation with a three-phase schedule that stabilizes query alignment before end-to-end refinement. On IU X-Ray, the model achieves BERTScore F1 of 0.700 and abnormal-study recall of 0.885, demonstrating strong semantic fidelity and clinically relevant abnormality sensitivity. Despite modest condition-specific F1 for low-prevalence findings, the framework provides a reproducible, clinically grounded, and data-efficient foundation for CXR report generation, with clear directions for rare-condition calibration and radiologist-centered evaluation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_018.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=zMRiNtnDAd
BibTex
@InProceedings{WagSat_CheXQuery_MICCAISAT2026,
author = { Wagle, Satish AND Jones, Clifford AND Baral, Saroj AND Poudel, Khem},
title = { { CheXQuery: Anatomy and Condition Guided Radiology Report Generation under Limited Data } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}