Abstract
Automatic radiology report generation from 3D computed tomography (CT) remains challenging because generated reports may contain unsupported findings or omit clinically relevant abnormalities. We propose an evidence-aware retrieval-augmented framework that retrieves similar training cases with a 3D CT-text contrastive model and converts retrieved reports into compact positive and negative finding checklists. These checklists are presented as auxiliary evidence to be verified against the current CT volume rather than as text to copy. We further apply RAG-aware supervised fine-tuning so that the generator learns to use this evidence during report generation. Evaluation on the CT-RATE validation set using the official RadBERT-based 18-label clinical efficacy protocol shows that prompt-only RAG consistently reduces hallucinated labels but increases omissions, resulting in significant CE macro-F1 degradation across all four retrieval strategies (all p < 0.001). The RAG-aware fine-tuned variants show partial numerical recovery from this degradation. Fusion-based SFT shows no statistically detectable difference in CE macro- or micro-F1 from the no-RAG baseline while significantly reducing hallucinated labels, whereas I2I achieves the highest observed BLEU and RadGraph ER F1 but retains a CE-F1 gap. Overall, the results reveal an over-conservative failure mode of prompt-only retrieval conditioning and highlight the need to jointly evaluate hallucination, omission, and clinical efficacy when incorporating retrieved evidence into 3D CT report generation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_013.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{HenYue_Balancing_MICCAISAT2026,
author = { Heng, Yue AND Hayashi, Yuichiro AND Oda, Masahiro AND Mori, Kensaku},
title = { { Balancing Retrieved Evidence for 3D CT Report Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}