Abstract
The HEADLINE challenge asks for a free-text radiology report to be generated from a whole non-contrast head CT volume, and ranks submissions by a composite text metric dominated by BERTScore. We describe a deliberately compact system. Each volume is resampled to an isotropic grid, reduced to a small budget of axial slices selected by foreground mass, and rendered so that three clinically motivated Hounsfield windows (brain, stroke, bone) occupy the three colour channels of every slice. A 3B-parameter general-purpose vision-language backbone is loaded in 4-bit precision and adapted with low-rank adapters on the language-model attention projections only, training 7.4M parameters, or 0.20% of the network. On a patient-disjoint internal split of 681 studies the system reaches a composite score of 0.6536 under our own metric implementation without BERTScore baseline rescaling (0.1794 with it, the convention used for official scoring), and on the official Validation Phase leaderboard it reaches 0.2461, placing second. We report three negative results we consider as informative as the positive one: extending the adapters to every linear layer, including the vision tower, did not improve validation score; beam search scored below greedy decoding; and multi-planar rendering could not be evaluated because distributed training never became stable. We also quantify the effect of the baselinerescaling convention used for official BERTScore computation: it moves the reported composite by a factor close to three while leaving the ranking of every configuration we tested unchanged.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HEADLINE_003.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=ii0x4HxJwm
BibTex
@InProceedings{BaiUjj_LowRank_MICCAISAT2026,
author = { Baid, Ujjwal AND Baheti, Bhakti AND Madabhushi, Anant},
title = { { Low-Rank Adaptation of a 3B Vision-Language Model for Head-CT Report Generation: HEADLINE 2026 } },
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}
}