Abstract
Automated LI-RADS categorization from multiphase CT must accommodate variable phase availability, physical-scale variation, severe class imbalance, and distribution shift. We developed a sevenclass classifier for the AMPLIFAI challenge using multiphase CT and organizer-provided lesion masks. Affine-aware preprocessing generated lesion-centered local and contextual tensors containing CT, mask, signeddistance, and valid-support channels. A shared compact 3D Vision Transformer then encoded the available phase-scale views, while learned missingview tokens and a 70-feature scalar pathway supported phase-aware fusion. Two same-lineage final-fit neural states were averaged, with a confidencegated tubular specialist permitted only to redistribute probability between LR-M and LR-TIV. Development used 461 training cases and a repeatedly inspected 59-case development set. Its composite of 0.84 (QWK 0.86; SCR 0.72) is historical post-selection estimate, and all five development cases across LR-1–LR-3 were misclassified. After the operating recipe was frozen, final fitting used all 520 eligible cases. The 174-case private evaluation yielded 0.48 (QWK 0.47; SCR 0.52). At manuscript submission, 0.52 was the highest SCR displayed on the leaderboard, while QWK remained the principal performance challenge. Source-held-out experiments, rare-class failures, and the independent performance gap were consistent with selection optimism and distribution shift.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMPLIFAI_009.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=kB1L6aFSLQ
BibTex
@InProceedings{FuXin_DualScale_MICCAISAT2026,
author = { Fu, Xinyi AND Sheng, Ke},
title = { { Dual-Scale Phase-Aware 3D ViT with Bounded Tubular Rescue for LI-RADS Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17271},
month = {pending},
page = {pending}
}