Abstract
We propose a sliding-window Vision-Language Model (VLM) for scalable, geometry-preserving report generation from variable-length longitudinal 3D Computed Tomography (CT). Radiologists compare prior and current scans to assess progression, treatment response, and new abnormalities, but automating this remains challenging because serial scans vary in slice number, coverage, and acquisition. Existing methods often focus on single timepoints or use fixed-geometry encoding that distorts voxel-space geometry and weakens cross-visit correspondence. Instead, we extract local features from overlapping 3D sub-volumes, aggregate them into compact whole-volume representations while preserving geometry, and fuse previous- and current-visit tokens with textual instructions for decoding by a Large Language Model (LLM). On a large in-house Japanese longitudinal dataset, our method achieved the best performance among longitudinal baselines on most metrics (BLEU-1/2/3/4 0.380/0.298/0.246/0.206, METEOR 0.348, GREEN 0.273, RadGraph-F1 0.214). On the public CT-RATE benchmark, it surpassed SIGMA on GREEN (0.399) and BERTScore (0.876), while supporting 512x512 volumes with 32-180 slices at practical inference time. These results suggest that geometry-preserving sliding-window 3D encoding is a practical foundation for longitudinal CT report generation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ELAMI_022.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{NguKha_Sliding_MICCAISAT2026,
author = { Nguyen, Khang C. AND Wang, Cheng AND Komuro, Takanori AND Oda, Masahiro AND Mori, Kensaku},
title = { { Sliding window-based local feature extraction for findings reports generation from longitudinal 3D CT volumes } },
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}
}