Advancing predictive oncology: Integrating clinical and radiomic models to optimize transarterial chemoembolization outcomes in hepatocellular carcinoma

Abstract

This article discusses the innovative use of computed tomography radiomics combined with clinical factors to predict treatment response to first-line transarterial chemoembolization in hepatocellular carcinoma. Zhao et al developed a robust predictive model demonstrating high accuracy (area under the curve 0.92 in the training cohort) by integrating venous phase radiomic features with alpha-fetoprotein levels. This noninvasive approach enables early identification of patients unlikely to benefit from transarterial chemoembolization, allowing a timely transition to alternative therapies such as targeted agents or immunotherapy. Such precision strategies may improve clinical outcomes, optimize resource utilization, and increase survival in advanced hepatocellular carcinoma management. Future studies should emphasize external validation and broader clinical adoption.

Key Words: Hepatocellular carcinoma; Radiomics; Transarterial chemoembolization; Alpha-fetoprotein; Predictive modeling; Machine learning; Computed tomography; Texture analysis; Treatment response; Personalized oncology

Core Tip: Radiomic analysis of computed tomography images-particularly texture and shape features-combined with clinical biomarkers such as alpha-fetoprotein, enables accurate prediction of response to transarterial chemoembolization in hepatocellular carcinoma, with area under the curve values exceeding 0.90. These noninvasive models allow early identification of non-responders, support personalized treatment selection, and may improve outcomes through timely initiation of alternative therapies in liver cancer management.

TO THE EDITOR

Transarterial chemoembolization (TACE) remains the mainstay of treatment for intermediate-stage hepatocellular carcinoma (HCC), particularly for patients not eligible for curative interventions. Despite its widespread use, the response to TACE is variable, with up to half of the patients demonstrating poor outcomes[1]. Early identification of likely non-responders is critical to avoid unnecessary toxicity and to enable timely consideration of alternative systemic therapies. In a recent study, Zhao et al[2] developed a combined predictive model that integrates computed tomography (CT)-based radiomic features with clinical biomarkers, namely alpha-fetoprotein levels, to forecast treatment response to first TACE in patients with HCC. Their model achieved an area under the curve (AUC) of 0.92 in the training cohort and 0.815 in the validation cohort, demonstrating excellent discriminatory power[2].

Radiomics-the process of extracting quantitative features from medical imaging-offers a non-invasive, high-throughput methodology for capturing intratumoral heterogeneity. Zhao et al[2] focused on venous phase CT images, which were found to be more informative than arterial or combined-phase images. This phase was chosen due to its superior depiction of washout, a characteristic feature in HCC. The resulting nomogram, combining radiomic and clinical data, provides clinicians with a practical tool to estimate treatment response probabilities at the bedside. Importantly, it allows early stratification of patients who are unlikely to benefit from TACE, thereby optimizing therapy selection and potentially improving survival outcomes.

Supporting findings from other recent studies strengthen this direction. Vosshenrich et al[3] reported 88.9% accuracy (AUC 0.96) using entropy and mean of positive pixels as key features in a nested decision tree model. An et al[4] achieved an AUC of 0.947 by integrating alpha-fetoprotein with wavelet-derived radiomic features, including wavelet.LHL_ ngtdm_Contrast (a second-order texture feature that quantifies contrast between neighboring gray levels, indicating tumor heterogeneity), within a random forest model. Niu and He[5] developed a CT-based radiomics nomogram combining volumetric and texture features, reporting a C-index of 0.844 in the training cohort and 0.831 in the validation cohort for predicting TACE refractoriness. Morshid et al[6] used shape features and a random forest model to predict time to progression, achieving 74.2% accuracy and highlighting the prognostic value of non-texture features in HCC. Shi et al[7] recently conducted an external validation study demonstrating strong reproducibility of key radiomic features across multiple scanners and institutions, further supporting the robustness of radiomics-based prediction. Meanwhile, Tipaldi et al[8] explored radiomics in more simplified or lower-resource settings, focusing on the feasibility of integrating imaging-derived models into routine workflows without heavy computational demand. These models consistently highlight the predictive value of features such as entropy, skewness, gray-level co-occurrence matrix contrast, and shape descriptors, which together capture tumor heterogeneity and underlying biological behavior.

Identifying non-responders before treatment not only spares patients from ineffective procedures and associated risks but also facilitates earlier transition to systemic therapies like lenvatinib or atezolizumab-bevacizumab. These insights have the potential to significantly improve treatment decisions for patients with limited liver reserve or multifocal disease. However, despite their strong predictive performance, radiomics and machine learning models are often viewed as ‘black box’ systems due to limited interpretability. This may hinder clinician confidence and real-world adoption, particularly in settings where transparency and reproducibility are critical. While promising, the model by Zhao et al[2] and others has limitations. Most were developed using retrospective, single-center datasets and require external validation across more diverse populations. Standardization of radiomic feature extraction, especially regarding peritumoral segmentation, remains a significant challenge. Moreover, linking radiomic patterns to underlying molecular or histopathological phenotypes (radiogenomics) could enhance interpretability and clinical acceptance.

Future research should focus on prospective trials, multi-center validation, and integration with other omics data to develop robust, generalizable models. Additionally, the development of automated pipelines using deep learning may reduce variability and enhance reproducibility. Zhao et al’s work[2] represents a significant advancement in predictive oncology for HCC and exemplifies the growing utility of radiomics in precision medicine. As we move toward more individualized care, such tools can help bridge the gap between imaging, biology, and clinical decision-making.