PaperlayerView exact revision on arXiv

Source-linked AI summary

AMPLIFAI: A Multiphase CT Dataset for Benchmarking Clinical Reasoning in LI-RADS Assessment of Liver Lesions

Pranav Kulkarni, Nikhil Shah, Amritansh Suryavanshi, Jana Delfino, James Tonascia, Jade Wong-You-Cheong, Barton Lane, Joseph Chirico, Jeffrey D. Hirsch, Ang Li, Heng Huang, Florence X. Doo

arXiv:2608.14778v1cs.CVcs.LG

TL;DR

LI-RADS characterization on multiphase CT lacks large public datasets with granular feature annotations, limiting available resources for AI research. AMPLIFAI addresses this gap with a documented 590-case dataset containing LI-RADS labels and voxel-level masks for three major imaging features, supporting standardized research while remaining intended for non-commercial research rather than clinical use.

  • Problem

    Large, publicly available multiphase CT datasets with granular radiologist annotations of major LI-RADS features are lacking.

  • Method

    The paper constructs and documents AMPLIFAI by harmonizing four public datasets and obtaining expert LI-RADS labels and feature-level segmentations.

  • Results

    AMPLIFAI provides 590 multiphase CT cases with LI-RADS categories and voxel-level masks for APHE, washout, and enhancing capsule.

  • Takeaways & Limitations

    AMPLIFAI establishes a standardized, transparently documented benchmark for AI-based LI-RADS characterization research.

  • Takeaways & Limitations

    The dataset is restricted to personal, non-commercial research and has not been approved by the FDA for clinical use or diagnosis.

Abstract

from arXiv · show

Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality worldwide, with early detection improving survival from <20\% to >70\%. The standardized LI-RADS criteria establish a biopsy-free, fully imaging-based framework that can serve as a foundation for automating HCC diagnosis with artificial intelligence (AI). However, the lack of large, publicly available datasets with high-quality labels has limited the development of AI models for LI-RADS characterization. We introduce the \textbf{AMPLIFAI} dataset, the first public dataset of multiphase abdominal CT scans annotated with LI-RADS categories and segmented for three major LI-RADS features: arterial phase hyperenhancement, washout, and enhancing capsule. Following the \emph{Datasheets for Datasets} format, this paper details the dataset's composition, curation process, and annotation pipeline to facilitate transparent, reproducible research.

1 Introduction

HCC remains a major mortality burden, while LI-RADS provides a biopsy-free imaging framework for diagnosis and AI-assisted early detection. AMPLIFAI addresses limited public data availability with a standardized, expert-annotated multiphase abdominal CT benchmark.

  • Clinical motivation: Early detection can improve HCC survival from <20% up to >70%, and LI-RADS defines LR-5 as “definitely HCC” within a biopsy-free imaging pathway.LI-RADS relies on abdominal CT for diagnosis and provides a foundation for AI-based early detection.
  • Data gap: Multiphase CT characterization captures dynamic lesion enhancement but requires spatial registration, phase-specific annotation, and cross-phase feature analysis.These requirements make curation time-consuming and contribute to a lack of large, publicly available datasets with granular feature-level labels.
  • Dataset contribution: AMPLIFAI provides a standardized benchmark comprising 590 cases harmonized from four public datasets for AI-based characterization of HCC lesions.Annotations were produced by five board-certified radiologists and one resident and include LI-RADS categories, lesion size, and voxel-level masks for three major LI-RADS features.
  • Scope and limitation: The datasheet covers only publicly released training and validation data, excluding the held-out private institutional test set to preserve challenge integrity.The paper identifies the AMPLIFAI Challenge Proposal and Official Website as primary compilation sources.

2 AMPLIFAI Datasheet … 2.2 Composition

AMPLIFAI is a public dataset for automated HCC characterization in multiphase abdominal CT using LI-RADS, addressing the lack of granular radiologist-annotated data. Each instance represents one patient CT study evaluated under this framework.

  • 2.1 Motivation: AMPLIFAI targets automated characterization of HCC lesions in multiphase abdominal CT using standardized LI-RADS criteria.LI-RADS categories range from LR-1, definitely benign, to LR-5, definitely HCC.
  • 2.1 Motivation: LI-RADS evaluates lesion morphology and dynamic contrast enhancement across non-contrast, arterial, portal venous, and delayed CT phases.This temporal analysis captures enhancement patterns across successive time points after contrast injection.
  • 2.1 Motivation: APHE is the primary gating feature, defined by the lesion appearing brighter than surrounding healthy tissue during the early arterial phase.The enhancement is considered non-rim when it spreads internally rather than remaining peripheral.
  • 2.1 Motivation: LI-RADS assessment combines lesion size with non-peripheral washout and enhancing capsule to determine HCC probability when APHE is present.Washout appears as portal venous or delayed-phase darkening, while an enhancing capsule is a peripheral rim retaining contrast.
  • 2. AMPLIFAI Datasheet: AMPLIFAI addresses the absence of large public datasets with granular radiologist-annotated LI-RADS feature labels by providing voxel-level masks for APHE, washout, and enhancing capsule.The dataset is intended to support automation of the spatio-temporal reasoning required for HCC diagnosis.
  • 2.1 Motivation: The dataset was created by the Center for Applied AI at the University of Maryland Institute for Health Computing with researchers, faculty, and clinicians from affiliated Maryland institutions.Contributing institutions include the University of Maryland, College Park, the University of Maryland School of Medicine, and the University of Maryland Medical System.
  • 2.2 Composition: Each dataset instance represents a single multiphase abdominal CT study of a patient evaluated for HCC using standardized LI-RADS criteria.The instance type is therefore a patient-level CT study rather than an isolated image or lesion.

How many instances are there in total (of each type, if appropriate)?

AMPLIFAI contains 590 multiphase CT cases from 584 unique patients, sourced from four public datasets and divided into patient-disjoint training and validation sets. Each case includes CT volumes, annotations, and segmentation masks for a target liver lesion and its LI-RADS features.

  • Dataset size and composition: 590 cases from 584 unique patients were collected from TCGA-LIHC, WAW-TACE, HCC-TACE-SEG, and PLC-CECT.Most cases correspond to one patient; TCGA-LIHC includes up to three visits for five patients.
  • Instance contents: Each instance is a multiphase contrast-enhanced abdominal CT study containing one arterial and at least one late contrast phase.Late phases may be portal venous and/or delayed; non-contrast scans are included when available, and volumes and masks are provided in NIfTI format.
  • Instance contents: Each case contains a LI-RADS category, target-lesion size, labels for three major LI-RADS features, and voxel-level masks for the lesion and those features.The three features are non-rim APHE, non-peripheral washout, and enhancing capsule.

Any other comments? · 2.3 Collection Process

AMPLIFAI was curated from four public multiphase CT datasets and annotated through a structured, blinded workflow involving experienced radiologists and a resident. The process used custom secure software, validation procedures, defined collection periods, and retrospective-study oversight.

  • 2.3 Collection Process: Five board-certified radiologists and one experienced radiology resident annotated cases, with three attending annotators and two validators maintaining a consistent, blinded protocol.The team received workflow onboarding and attended regular check-in meetings.
  • 2.3 Collection Process: Each case was independently annotated by the resident and one attending annotator using prior lesion annotations and AI-generated feature masks, with nnU-Net predictions refined manually after five cases.The iterative workflow was intended to reduce annotation burden.
  • 2.3 Collection Process: Validators reviewed completed cases independently of initial annotation and could select, union, or intersect annotations, while disagreements could be returned to annotators for verification.This addressed equivocal features and differing target-lesion selections.
  • 2.3 Collection Process: Annotations were created in a custom 3D Slicer module hosted in a secure hospital research environment with authenticated remote access.The interface supported lesion and all three major LI-RADS feature segmentations on multiphase CT.
  • 2.3 Collection Process: An interdisciplinary University of Maryland team curated the dataset, drawing on researchers, faculty, and clinicians from affiliated academic and medical institutions.The annotation team included six clinicians with substantial abdominal-imaging experience.
  • 2.3 Collection Process: The dataset comprises multiphase CT studies from TCGA-LIHC, WAW-TACE, HCC-TACE-SEG, and PLC-CECT, whose source data were collected between November 2002 and December 2022.The annotations were performed between May 2026 and June 2026.
  • 2.3 Collection Process: The retrospective study received an IRB exemption from the University of Maryland Institutional Review Board under protocol HP-00117149.Patients were not directly notified because the data came from publicly available repositories.
  • 2.3 Collection Process: The data are anonymized without personally identifiable information, but patients cannot be removed from the dataset because it derives from publicly available repositories.The dataset contains abdominal CT scans from human patients.

Any other comments? · 2.4 Preprocessing/Cleaning/Labeling

The dataset underwent standardized preprocessing across four source datasets, including format harmonization, phase detection, multiphase registration, and registration-quality assessment. Raw data remain publicly accessible through source repositories, and the complete preprocessing code will be released for reproducibility.

  • 2.4 Preprocessing/Cleaning/Labeling: Preprocessing standardized imaging data across all four source datasets before downstream use.The pipeline included removal of incompatible series, format conversion, phase detection, registration, and quality assessment.
  • 2.4 Preprocessing/Cleaning/Labeling: Irrelevant DICOM and DICOM-SEG series were removed, including Scout and Dose Reports and combined lesion segmentations incompatible with per-lesion annotations.The exclusions addressed source-specific series that did not fit the dataset’s annotation schema.
  • 2.4 Preprocessing/Cleaning/Labeling: DICOM datasets were converted to NIfTI with dcm2niix, while WAW-TACE lesion segmentations were converted from NRRD to NIfTI for format consistency.The conversions standardized image volumes and segmentation masks across sources and improved accessibility for research use.
  • 2.4 Preprocessing/Cleaning/Labeling: Contrast phases were classified from imaging metadata or, when unavailable, predicted with Comp2Comp.The target cohort required an arterial phase and at least one late-contrast phase.
  • 2.4 Preprocessing/Cleaning/Labeling: All multiphase CT studies were registered with ANTsPy to a fixed reference, prioritizing the portal venous phase and using delayed phase when necessary.This established spatial correspondence across phases.
  • 2.4 Preprocessing/Cleaning/Labeling: A Dice score above 0.64 between rigid bone structures segmented with TotalSegmentator defined acceptable registration quality.This criterion was used to identify poorly registered pairs.
  • Any other comments?: Raw data remain publicly accessible through TCIA, Zenodo, and ScienceDB, while the complete preprocessing code will be made publicly available.TCGA-LIHC and HCC-TAGE-SEG are hosted on TCIA, WAW-TACE on Zenodo, and PLC-CECT on ScienceDB; the code release is intended to support transparency and reproducibility.

Any other comments? · 2.5 Uses · What (other) tasks could the dataset be used for?

AMPLIFAI supports LI-RADS-based HCC lesion characterization and broader medical foundation-model research, while its use is restricted to personal, non-commercial research and not clinical diagnosis or individual identification. A partial list of papers using the dataset will be provided on the AMPLIFAI Challenge Website.

  • 2.5 Uses: The dataset has been used to train models that characterize HCC lesions according to standardized LI-RADS criteria.These models use multiphase abdominal CT scans and a target lesion mask to predict the probability of an associated LI-RADS category.
  • 2.5 Uses: The models predict probabilities for LI-RADS categories LR-1 through LR-5, LR-TIV, or LR-M.One explored approach strictly follows clinical criteria to predict major LI-RADS features before categorization.
  • Any other comments?: A partial list of papers using the dataset will be included on the AMPLIFAI Challenge Website.The passage identifies the website as an access point for papers or systems using the dataset.
  • What (other) tasks could the dataset be used for?: Beyond LI-RADS characterization, the dataset could support medical foundation models for visual-question answering and reasoning.The passage identifies both visual-question answering and reasoning as potential applications.
  • What (other) tasks could the dataset be used for?: The dataset should be used only for personal, non-commercial research.This restriction is stated as a condition on dataset use.
  • What (other) tasks could the dataset be used for?: The dataset has neither been reviewed nor approved by the U.S. Food and Drug Administration for clinical use.The passage specifically states that FDA review or approval has not occurred.
  • What (other) tasks could the dataset be used for?: The dataset should not be used to identify individuals or diagnose patients with pathologies.These prohibited uses are stated alongside the restriction against clinical use.

Any other comments? … Any other comments?

AMPLIFAI is currently publicly available through the challenge website and Hugging Face for challenge participants. It includes images, labels, and metadata for training and validation splits, totals about 146 GB, and is licensed for personal, non-commercial research use under CC BY-NC-SA.

  • Any other comments?: The distribution comments indicate release to third parties through the AMPLIFAI Challenge Website and Hugging Face.
  • 2.6 Distribution: The dataset is publicly available on the AMPLIFAI Challenge Website.
  • 2.6 Distribution: Challenge participants receive the dataset through Hugging Face.
  • 2.6 Distribution: The distribution contains all images, labels, and metadata for the training and validation splits.
  • 2.6 Distribution: 146 GB is the approximate total dataset size.
  • When will the dataset be distributed?: The dataset is currently available.
  • When will the dataset be distributed?: CC BY-NC-SA licenses the dataset for personal, non-commercial research use.
  • When will the dataset be distributed?: No export controls or other regulatory restrictions apply because all data is sourced from publicly available repositories.

2.7 Maintenance … Any other comments?

The dataset will be maintained by the University of Maryland Institute for Health Computing, with no current plans for updates and no limits on information retention. Users may extend or build on the dataset for non-commercial research with attribution under CC BY-NC-SA, but no contribution mechanism is provided.

  • Who will be supporting/hosting/maintaining the dataset?: The University of Maryland Institute for Health Computing will maintain the dataset.
  • How can the owner/curator/manager of the dataset be contacted (e.g., email address)?: The passages provide no answer about whether an erratum exists or how the dataset owner can be contacted.
  • How can the owner/curator/manager of the dataset be contacted (e.g., email address)?: There are no plans to update the dataset at this time.
  • How can the owner/curator/manager of the dataset be contacted (e.g., email address)?: There are no limits on retention of the information in the dataset.
  • Any other comments?: The passages do not specify whether older dataset versions will remain supported, hosted, or maintained.
  • How can the owner/curator/manager of the dataset be contacted (e.g., email address)?: No mechanism is provided for others to contribute extensions or augmentations to the dataset.
  • Any other comments?: Users are encouraged to extend, augment, or build on the dataset for non-commercial research use with attribution under CC BY-NC-SA.
Loading 2608.14778v1…