Data
Challenge Data
The challenge dataset comprises 779 multi-phase, contrast-enhanced CT studies, divided into training, validation, and test sets. A case refers to a single patient containing one arterial and at least one late contrast phase (either venous or delayed). For every case, the multi-phase CT volumes are provided in NIfTI format.
In addition, each case includes exactly one lesion segmentation, also in NIfTI format. This segmentation is a voxel-level mask of the target lesion annotated by board-certified radiologists. When present, voxel-level segmentations of key imaging features - non-rim APHE, non-peripheral washout, and enhancing capsule - are also provided in NIfTI format. While multiple lesions may be present within a given study, only one lesion per case is annotated and released. This annotated lesion represents the most clinically relevant or prominent lesion, as selected by the radiologists.
Participants should develop models that predict the LI-RADS category for the single annotated lesion only, even if other lesions are visible in the images.
Competition bundles are distributed through the Codabench competition page. For access issues, email amplifai@som.umaryland.edu.
Dataset Summary
| Split | Cases |
|---|---|
| Training set | 531 |
| Validation set | 59 |
| Test set (private) | 189 |
Training and Validation Data
The public training and validation dataset was curated from four publicly accessible datasets that were harmonized and preprocessed for this challenge, then re-annotated according to the challenge protocol. The contributing datasets and their source institutions are listed below:
- WAW-TACE - Medical University of Warsaw
- HCC-TACE-SEG - The University of Texas MD Anderson Cancer Center
- TCGA-LIHC - Mayo Clinic, Rochester; University of North Carolina, Chapel Hill; Alberta Health Services; Lahey Hospital & Medical Center
- PLC-CECT - Chongqing Yubei District People’s Hospital
Cases from contributing datasets that included existing annotations were partially pre-annotated; all training and validation cases were fully re-annotated by three board-certified radiologists following the challenge protocol.
The validation set comprises approximately 10% of the public data and is constructed to ensure equitable representation across all contributing datasets.
Test Data
The held-out test set is not public and was specifically curated for this challenge. The test set contains 189 cases acquired between February 2023 and May 2026 within three different imaging centers of the University of Maryland Medical System. For each test case, participant models receive the multi-phase CT volumes and a lesion segmentation for the intended target lesion as input. The test set is private and will not be released publicly; to ensure the integrity of the challenge, no additional information is provided.
Evaluation
Models will be evaluated solely on the accuracy of LI-RADS categorization.
The final ranking score will be computed as a weighted composite metric that reflects both the primary ordinal LI-RADS characterization task and recognition of special LI-RADS categories for non-HCC lesions. Each algorithm is required to generate a single LI-RADS category for the pre-defined target lesion in each test case. Valid prediction labels are LR-1, LR-2, LR-3, LR-4, LR-5, LR-M, and LR-TIV.