Abstract
The Liver Imaging Reporting and Data System (LI-RADS) provides a rule-based framework for categorizing liver lesions in at-risk patients based on clinically defined imaging features, including lesion size, arterial phase hyperenhancement, washout, capsule appearance, and management-relevant special findings. Learning LI-RADS categories directly from a small and severely imbalanced multiphase CT cohort requires a model to simultaneously discover these imaging findings and infer their clinical composition rules, posing substantial challenges in data efficiency and interpretability. We propose Read-to-Decide (R2D), a clinically structured two-stage framework that aligns model inference with the semantic representation underlying LI-RADS. In the first stage, the model extracts LI-RADS major features from multiphase imaging and represents them as probabilistic predictions. In the second stage, these feature probabilities are combined with lesion geometry by a tree ensemble to predict the final LI-RADS category. Because imaging evidence enters the final decision through explicit feature probabilities, the frozen model also supports direct test-time intervention: replacing an erroneous feature prediction with a radiologist’s reading can correct the final category, making errors both attributable and actionable. On the AMPLIFAI cohort, comprising 461 training cases and 59 validation cases, our method achieves an official validation score of 0.676 and an out-of-distribution test score of 0.51, ranking first on the validation leaderboard and second on the updated test leaderboard.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AMPLIFAI_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=3Swr41l81o
BibTex
@InProceedings{LiuYua_Read_MICCAISAT2026,
author = { Liu, Yuanye AND Zhuang, Xiahai},
title = { { Read to Decide: Interpretable Feature-Guided LI-RADS Categorization from Multiphase CT } },
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}
}
