WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data

Introduction

Hepatocellular carcinoma (HCC) is the most prevalent primary liver malignancy, encompassing a heterogeneous group of patients with varying prognoses (1). Managing HCC includes various therapeutic options, such as liver transplantation, liver resection, local-regional therapies, chemotherapy and immunotherapy (2). Currently, transarterial chemoembolization (TACE) is the treatment of choice for patients with unresectable HCC without extrahepatic spread and patent venous vessels. Artificial intelligence (AI) research in HCC is evolving rapidly, including use of models to enhance HCC detection, segmentation, and prognostication. Development of such models requires comprehensive data, including imaging and clinical variables (3–5). However, the majority of these studies use internal datasets that are not publicly available alongside the manuscripts and thus lack external data for model validation or performance tuning (6–8). Moreover, existing publicly available datasets have a limited number of cases with imaging and clinical data or tumor segmentations, which are crucial for AI research purposes (9,10). Substantial efforts are being made to provide these data to enhance HCC research.

Here, we present the annotated WAW-TACE dataset, which includes data from patients with HCC treated with TACE at the Medical University of Warsaw (hereafter, MUW), offering an important data resource for diverse researchers. This dataset consists of extensive baseline clinical data, pre-TACE multiphase CT imaging with segmentation masks of HCC lesions and multiple anatomic structures, radiologic therapy response assessments, and crucial patient time-dependent outcome measures such as overall survival (OS), progression-free survival (PFS), and time to favorable treatment response.

Materials and Methods

This retrospective dataset includes 233 treatment-naive patients diagnosed with unresectable HCC and treated with TACE at the Central Clinical Hospital of the MUW between May 2016 and April 2021. The study was approved by the Bioethics Committee at the MUW (decision no. AKBE/41/2024). Figure 1 provides an overview of patient inclusion criteria, the size of the cohort, and the components of the dataset.

Figure 1:

Download as PowerPoint

The inclusion criteria were: (a) at least one HCC lesion; (b) patients receiving conventional TACE treatment; (c) satisfactory liver function (Child-Pugh class A or B) with no evidence of metastatic disease; and (d) the availability of contrast-enhanced CT prior to the first TACE treatment.

Exclusion criteria were: (a) patients undergoing liver transplantation, resection, ablation, or drug-eluting beads TACE before or during the follow-up period; (b) history of malignant neoplasms other than HCC; (c) metastatic disease; (d) large vessel invasion; and (e) patients with known uncontrolled functional or metabolic diseases that could potentially bias survival (such as myocardial infarction or stroke).

CT Imaging and Data Processing

CT protocol.— All patients underwent multiphasic CT within 90 days before the initial TACE cycle. A detailed list of information regarding CT scanners, acquisition parameters, and available examination phases is presented in Table 1.

Table 1: Details of CT Imaging Phases, Scanners, and Acquisition Parameters

Volume of interest segmentation.— The exact section thicknesses for each individual study is reported in a separate column in the accompanying ct_hcc_metadata file. Each CT examination underwent de-identification and was converted to NIfTI format. No additional filtering or CT image normalization was performed.

The segmentation masks for internal organs were generated using the nnU-Net deep learning model TotalSegmentator (15,16). This algorithm enables the unsupervised segmentation of volumes of interest (VOIs), such as the liver, spleen, and kidneys, among others. For all segmentation tasks, we utilized the pretrained model. No additional training was conducted, as we employed the default hyperparameters as previously described by Wasserthal et al (16). In our dataset, native, arterial, portal, and delayed phases were separately segmented, each with 104 VOIs.

Additionally, for the segmentation of HCC lesions, the TotalSegmentator extension was employed. These VOIs were manually corrected in Slicer 3D software (version 5.4.0) by a radiologist (K.B., with 6 years of overall experience, including 2 years specifically in medical image segmentation) (17). Two independent radiologists conducted validations and made the necessary modifications (K.K. and K.L., with 14 and 11 years of experience in TACE and abdominal radiology, respectively). All observers adhered to a uniform protocol for HCC segmentation, which involved outlining the tumor boundaries on subsequent axial CT sections. These were then connected using the “fill between sections” option with subsequent mask’s smoothing using a 3-mm Gaussian filter.

Feature extraction.— We used PyRadiomics (version 3.0.1) to extract features categorized into classes, ensuring adherence to the Imaging Biomarker Standardization Initiative (18). Importantly, all radiomics features were extracted from raw masks generated by TotalSegmentator, without manual VOI correction. For each VOI in the native, late arterial, portal venous, and delayed CT examination phases, there were 32 features belonging to two groups: shape-based features (n = 14) and first-order statistics (n = 18). Furthermore, from the largest liver tumor VOI, an additional 75 gray-level radiomics features were extracted, including 24 gray-level co-occurrence matrix; 14 gray-level dependence matrix; 16 gray-level size zone matrix; 16 gray-level run length matrix; and five neighboring gray-tone difference matrix.

Results

The WAT-TACE dataset can be downloaded (https://zenodo.org/records/12741586) and includes four major components, as outlined in Figure 1: clinical data, imaging, segmentations, and radiomics features.

Clinical Data

Clinical variables, including outcome measures for the patients, are presented in tabular format in a separate file (“clinical_data”). Table S1 presents all headers from the clinical data spreadsheet and variables’ definitions. The baseline characteristics of the 233 patients (median age, 66 years (range: 28–86 years); 185 male, 48 female) included in the dataset are summarized in Table 2. The median OS was 27.3 months, with 65 censored patients (27.9%) and a minimum follow-up duration of 33.8 months for surviving patients. The median PFS was 17.7 months, with 60 observed events and 55 censored patients.

Table 2: Baseline Characteristics of Patients Included in the WAW-TACE Dataset

Segmentations

The WAW-TACE dataset collectively contains 377 HCC lesion segmentations and 104 organ VOIs per patient, available in files “tumor_masks” and “organ_masks,” respectively. A representative set of segmented images from a single WAW-TACE patient is presented in Figure 2. Additional examples of segmented images are presented in Figure 3.

Figure 2:

Download as PowerPoint

Figure 3:

Download as PowerPoint

Discussion

In the expanding era of AI-based research in oncologic imaging, there is a growing demand for extensive data to facilitate the development of prediction models and enhance their generalizability. The WAW-TACE dataset substantially expands the number of publicly accessible annotated HCC lesions imaged using CT, facilitating clinical as well as AI research aimed at predicting time-dependent treatment outcomes. The comprehensive nature of this dataset positions it as an important resource for various researchers.

Our collection adds a substantial amount of annotated data to another recently published TACE dataset (9). In comparison, we also include patients with multiple HCC lesions, not just single tumors. This approach reflects more closely the clinical context of TACE treatments. Moreover, the inclusion of multiphase CT examinations, multiple organ segmentation masks, radiomics features, and comprehensive clinical data combined with time-dependent outcome measures provides a new opportunity for researchers to explore the potential use of AI models. The dataset facilitates robust model training by providing a large, diverse collection of annotated HCC lesions, which is required for developing AI models that can accurately detect and segment tumors. These models can then be used to improve clinical decision-making by providing precise and reliable predictions about disease progression, response to treatments such as TACE, or survival. For example, a part of this dataset was used in a recent study where we used radiomics for TACE outcome prediction, relying solely on automated segmentations (19). Furthermore, the WAW-TACE dataset allows for interinstitutional comparisons and cross-validation, enabling researchers from various institutions to evaluate and test their models against current standards. This can lead to enhanced model performance and generalizability, as models trained and tested on diverse datasets are more likely to perform well across patient populations and imaging conditions. In addition, the manually created HCC segmentation masks can be used to investigate and tailor AI algorithms for tumor segmentation and classification. The dataset reflects the complexity of clinical scenarios by including a variety of cases with multiple lesions, allowing for the development of AI tools capable of addressing real-world challenges in oncologic imaging. Finally, the dataset’s extensive clinical data can aid research into personalized medicine. The models trained on this dataset can be used to predict individual patient outcomes by optimizing treatment plans based on tumor characteristics and patient profiles.

There are several limitations worth mentioning. First, only baseline imaging is available in the dataset. Informative image features that may appear on postprocedural scans, or through the assessment of changes over time, cannot be evaluated with this dataset. Second, not all patients included have all CT phase series available, leading to some missing image data. Furthermore, the imaging data comes from different CT systems and the median section thicknesses varied between patients, which should be emphasized when using the data for radiomics analysis. Notably, no specific CT image preprocessing or resampling was performed prior to radiomics feature extraction. Finally, the most reliable study end point was OS, recorded objectively using the national government database. However, TACE treatment response assessments and PFS were more subjective, relying on the reevaluation of post-TACE imaging by one author (K.B.). The process involved using the LI-RADS treatment response algorithm, cross-referencing imaging with clinical data, radiology reports, and laboratory data. Despite these measures, the subjective nature of response evaluation may introduce some variability.

In summary, this interdisciplinary resource is an important contribution to HCC research and facilitates use of AI in medical imaging research aimed at enhancing patient outcomes.

Keywords: Liver, Oncology, Hepatocellular Carcinoma