Abstract
Non-contrast head CT is central to the rapid assessment of neurological emergencies and other clinically significant intracranial abnormalities, including hemorrhage, ischemia, traumatic injury, mass effect, hydrocephalus, and tumors. Growing examination volumes place increasing pressure on timely and consistent radiology reporting. To advance automated reporting, the MICCAI 2026 HEADLINE challenge represents a dataset and a benchmark for developing and evaluating vision–language models that generate free-text radiology reports from 3D non-contrast head CT volumes. To this end, we developed FiReGen as a hybrid framework combining findings-aware retrieval, vision-language generation, and consensus selection, based on 16,468 studies from 10,488 patients, each paired with a radiologist-authored report. On the official 500-study validation set, the retrieval-only submission achieved a composite score of 0.2400, comprising a BERTScore of 0.2634, ROUGE-L of 0.2800, METEOR of 0.1697, and BLEU-4 of 0.1064. These results demonstrate that findings-aware retrieval provides a promising foundation for automated 3D head CT reporting, while the reliable integration and selection of generated reports remain important areas for improvement.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HEADLINE_002.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=CMZQW8diWf
BibTex
@InProceedings{KacSan_FiReGen_MICCAISAT2026,
author = { Kachole, Sanket AND Bakas, Spyridon},
title = { { FiReGen: Findings-Aware Retrieval and Reinforcement-Tuned Generation of Free-Text Head CT Radiology Reports } },
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}
}
