Radiomics-Based Delineation of Gross Tumor Volume and Normal Liver Tissue in Hepatocellular Carcinoma using Contrast-Enhanced CT and Ensemble Machine Learning

introduction: 1. Introduction Primary liver cancer remains one of the leading causes of cancer-related mortality worldwide, with hepatocellular carcinoma (HCC) accounting for approximately 75%–85% of cases. The incidence of HCC has continued to rise in recent years, representing a significant global health burden.[1,2] Diagnosis of HCC typically relies on a combination of imaging findings and serological assessment, while treatment selection is guided by tumor staging, residual liver function, and the overall condition of the patient.[3] For those ineligible for surgical resection, alternatives treatments include radiofrequency ablation, transarterial chemoembolization (TACE), molecular targeted therapy, immunotherapy, and radiotherapy (RT), applied with either palliative or curative intent.[4] The increasing implementation of image-guided radiotherapy, intensity-modulated radiotherapy, and stereotactic body radiotherapy have positioned RT as a viable modality for local tumor control and prolonging survival outcomes.[5,6] Accurate delineation of the target volume is fundamental in radiotherapy to ensure adequate dose delivery while minimizing exposure to uninvolved tissues.[7] In accordance with the recommendations of the International Commission on Radiation Units and Measurements Report 83, radiotherapy planning involves the sequential definition of the gross tumor volume (GTV), clinical target volume (CTV), and planning target volume (PTV).[8] The GTV, as the fundamental representation of the lesion entity, serves as the central reference for dose–volume histogram analysis and provides the geometric foundation for expansion to the CTV and PTV. Underestimation may result in omission of microscopic disease and marginal recurrence, whereas overestimation may irradiate normal hepatic parenchyma, increasing the risk of radiation-induced liver disease.[9,10] Accurate differentiation between malignant and non-malignant hepatic tissue is therefore essential for delivering individualized and safe RT. In clinical practice, however, GTV delineation remains largely dependent on the subjective judgment of radiation oncologists. In HCC imaging, coexisting hepatic conditions such as cirrhosis, steatosis, and regenerative nodules, combined with poorly defined tumor–parenchyma boundaries, contribute to considerable inter-observer and intra-observer variability. This subjectivity reduces both the precision and reproducibility of radiotherapy planning.[11] Furthermore, manual contouring is time-consuming and labor-intensive, particularly in cases involving multiple lesions or large tumor volumes. During radiotherapy, dynamic changes in tumor size and hepatic position resulting from respiration and physiological variations add further complexity to re-delineation of the target.[12] These challenges underscore the need for automated, objective, and standardized mechanisms for GTV identification to enhance delineation efficiency, minimize human error, and support the implementation of adaptive radiotherapy strategies in the management of HCC.[13] Radiomics, a quantitative imaging approach that extracts latent biological information from medical images, has been increasingly applied across diverse oncological domains, including tumor classification, prognostic assessment, treatment response prediction, and delineation of target volumes.[14] The process typically involves standardized acquisition and processing of imaging modalities such as computed tomography (CT) and magnetic resonance imaging (MRI), followed by high-throughput extraction of quantitative features related to texture, grayscale intensity, and morphology. These features are subsequently analyzed using artificial intelligence–based methods to construct predictive or diagnostic models. Radiomics thereby enables non-invasive, objective evaluation of intratumoral heterogeneity, offering considerable clinical interpretability and scalability. In the context of HCC, radiomics has been employed to predict tumor grade, assess microvascular invasion, and evaluate treatment response to TACE, demonstrating both clinical feasibility and potential for broader integration into clinical practice.[15–17] However, while substantial research has focused on radiomics-based classification and prognostic assessment in HCC, its application in the automated differentiation of tumor tissue from normal hepatic parenchyma within the radiotherapy workflow remains limited. Limited emphasis has been placed on improving the specificity of target volume identification. Advancing research in this field is essential not only for protecting normal liver tissue and mitigating radiation-related complications but also for optimizing individualized dose distribution and supporting adaptive radiotherapy strategies, thereby establishing it as a critical current research priority.[18] The present study utilized contrast-enhanced CT imaging to perform radiomics analysis of GTV regions and normal liver tissue regions in patients diagnosed with HCC. Multiple widely applied machine learning classification models were developed, and their discriminative performance and potential clinical applicability in differentiating tumor tissue from non-tumor tissue were systematically evaluated. The objective was to establish an interpretable, standardized, and generalizable auxiliary tool for radiotherapy planning, thereby providing methodological evidence and technical support for the automation and intelligent integration of target volume delineation within HCC radiotherapy workflows. materials and methods: 2. Materials and Methods 2.1 Patient Population This retrospective study included patients who underwent preoperative contrast-enhanced CT between December 2019 and December 2024, had histopathologically confirmed HCC, and possessed complete clinical records. Inclusion criteria were (1) availability of at least one preoperative contrast-enhanced CT scan including both arterial and portal venous phases; (2) no prior history of radiotherapy for liver-related malignancies or systemic chemotherapy; (3) histopathological confirmation of HCC through surgical resection or biopsy; (4) estimated survival time of at least three months and absence of definitive contraindications to radiotherapy; and (5) adequate image quality for subsequent image processing and radiomics analysis. Exclusion criteria were (1) presence of severe motion artifacts, metallic artifacts, or image occlusion; (2) maximum tumor diameter less than 1.0 cm, precluding reliable region of interest (ROI) delineation; (3) concomitant presence of other primary or metastatic liver tumors; and (4) incomplete clinical information or unavailability of relevant CT image sequences. A total of 82 patients met the eligibility criteria. Among these, some presented with multiple independent lesions, yielding 91 tumor regions. Each tumor region was paired with a corresponding normal liver tissue region to serve as a control. These image sets were subsequently used for ROI delineation, radiomic feature extraction, and model development. To minimize the potential correlation bias associated with multiple lesions from the same patient, a cross-validation strategy was applied in the subsequent modeling process to enhance model generalizability. 2.2 Image Acquisition and Processing All imaging data were obtained using multi-detector spiral CT scanners, specifically Siemens SOMATOM Definition Flash, Canon Aquilion ONE (TSX series), and Philips Brilliance 64 systems. Scans were performed according to a standardized protocol with uniform parameter settings: pitch of 1.0, field of view of 35 cm, tube voltage of 120 kV, and slice thickness of 5 mm. Contrast-enhanced scans were performed using non-ionic iodinated contrast agents (iodine concentration: 270–350 mg I/mL) administered through the antecubital vein at a rate of 3.0 mL/s. The total contrast volume was calculated according to body weight (1.5 mL/kg). Scanning was conducted sequentially in three phases: arterial phase at 25 seconds, portal venous phase at 60 seconds, and delayed phase at 180 seconds. To reduce variability associated with scanner type and reconstruction parameters, all CT datasets underwent standardized preprocessing. Voxel size was resampled for consistency, and Gaussian filtering was used for image smoothing to enhance the reproducibility and stability of subsequent radiomic feature extraction. 2.3 ROI Segmentation ROI segmentation was conducted on portal venous phase images of contrast-enhanced CT using 3D Slicer software. For each patient, two ROI regions were delineated: (1) the GTV corresponding to the HCC and (2) the normal liver tissue region distant from the tumor. All GTV delineations were independently performed by two researchers with extensive experience in CT image interpretation and were subsequently reviewed by a senior radiologist on a case-by-case basis to ensure accuracy and consistency of annotations. The normal liver tissue ROIs were selected from the contralateral liver lobe that appeared radiologically normal and maintained an intact anatomical structure. During the delineation process, areas suspected of containing lesions, liver margins, and intrahepatic vascular or biliary structures were excluded to minimize feature variability and to enhance the robustness and representativeness of the extracted radiomic features. 2.4 Feature Extraction and Selection A total of 1,209 high-dimensional radiomic features were extracted from the delineated GTV and contralateral normal liver tissue regions using the PyRadiomics software package. Extracted features included grey-level histogram metrics, first-order statistical parameters, and texture features derived from