Abstract
Automated reporting of three-dimensional head computed tomography (CT) must combine volumetric image processing with free- text generation under practical inference constraints. We developed a lightweight dual-route system for the HEADLINE MICCAI 2026 chal- lenge. Each soft-tissue and bone reconstruction was reduced to eight evenly spaced axial slices and encoded with a frozen CLIP ViT-B/16. A trainable aggregator with T5-small generated one candidate report, while a second route predicted one of 16 report archetypes and returned its representative medoid report. A gradient-boosted selector chose be- tween the candidates per study. On an internal patient-separated eval- uation set of 1,711 studies, the Linux container achieved an unrescaled local composite proxy of Ssent local = 0.5680590, compared with 0.5628781 for T5 alone and 0.5616516 for the representative route. In a separate auxiliary analysis confined to the development set, we probed the same frozen features using report-derived weak labels. Attention aggregation increased midline-shift AUROC from 0.8006 to 0.8403 and mass-effect AUROC from 0.7169 to 0.7854 in all five folds, but not intracranial- haemorrhage AUROC (0.7763 to 0.7616). Adding bone-stream features provided no measurable incremental value for these three findings un- der the evaluated design. Finding–report disagreement did not reliably identify report errors, so the probe was not integrated. The results es- tablish an internal report-similarity improvement for candidate selection and delimit the auxiliary probe to representation analysis rather than clinical diagnosis or report quality control.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/HEADLINE_006.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=mK7GbDbpYl
BibTex
@InProceedings{YosSho_Head_MICCAISAT2026,
author = { Yoshimoto, Shohei},
title = { { Head CT Report Generation from Subsampled Slices with Candidate Selection } },
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}
}
