1,200 contrast-enhanced abdominal CT volumes for liver lesion segmentation and RECIST measurement

Overview

We are constructing a benchmark dataset for automated liver lesion detection and volumetric segmentation on contrast-enhanced CT of the abdomen, targeting hepatocellular carcinoma (HCC), colorectal liver metastases (CRLM), and benign focal liver lesions (hemangioma, cysts, FNH). Imaging protocol must include arterial-phase and portal-venous-phase acquisitions, with slice thickness ≤2 mm and pixel spacing ≤0.8 mm in-plane. The Hounsfield unit dynamic range of interest is –200 to +300 HU (liver parenchyma typically 50–70 HU in portal phase; hypervascular HCC peaks at 80–120 HU in arterial phase). Multiplanar reconstruction (MPR) in coronal and sagittal planes is welcome but not mandatory. Annotation requirements are demanding: each lesion must have a 3D segmentation mask generated or confirmed by an abdominal radiologist with at least 5 years of subspecialty experience. Masks should be provided in NIfTI or NIfTI-2 format, with a JSON metadata file encoding lesion type, RECIST 1.1 longest axial diameter in millimeters, lesion number, LI-RADS category for HCC cases, and whether the patient had prior locoregional therapy (TACE, ablation). Liver parenchyma whole-organ masks are strongly encouraged as an additional annotation layer to facilitate liver-volume normalization during training. De-identification must satisfy HIPAA Safe Harbor (Method 1) with removal of all 18 identifiers and re-mapping of DICOM UIDs. Cases with prior abdominal surgery or transplant should be flagged in metadata. We require a minimum of 20% negative cases (no focal lesion) to anchor the model's specificity. Tube voltage should be documented for each acquisition phase, with standard portal-venous-phase protocols at 100–120 kVp using automated tube-current modulation. Contrast agent type (iodinated, concentration in mg/mL), injection rate (mL/s), and delay time (seconds from injection to scan start) must be recorded in the JSON sidecar for each phase, as these parameters directly affect lesion-to-liver contrast and Hounsfield unit values at the time of acquisition. Scanner heterogeneity across multiple vendors (GE, Siemens, Philips) and field-site geographic diversity are required to prevent model overfitting to a single institution's acquisition style. QA exclusion criteria include studies with gross motion artifact, incomplete hepatic coverage, or absence of a portal-venous phase. Inter-rater segmentation agreement (Dice ≥ 0.80 on lesions ≥10 mm) must be documented per contributing site. The dataset will be used to develop a clinical decision-support tool for oncology multidisciplinary tumor boards, enabling automated lesion tracking across treatment cycles in compliance with RECIST 1.1 response assessment criteria. Data will be processed within an ISO 27001–certified cloud environment under a fully executed DUA.

Progress

Data Specifications

Use Cases

• Training and validating Medical imaging AI/ML models
• Benchmarking Medical imaging detection and segmentation algorithms
• Building de-identified Medical imaging research datasets for academic studies
• Augmenting existing Medical imaging datasets to reduce class imbalance