• Hepatocellular Carcinoma
• Immunotherapy

We have read with great interest the recent publication entitled “Pretreatment CT-based Machine Learning Radiomics Model Predicts Response in Unresectable Hepatocellular Carcinoma Treated with Lenvatinib plus PD-1 Inhibitors and Interventional Therapy” by Hua et al.1 Although lenvatinib combined with programmed cell death protein-1 inhibitors and interventional (LPI) therapy can be used as primary treatment options, some patients may still be non-responsive to LPI therapy and experience significant adverse events. However, there are no reliable and accurate biomarkers to predict a patient’s response to LPI therapy. To this end, this study developed and validated a radiomics-based machine learning model to non-invasively predict the efficacy of this triple therapy, and to perform risk stratification in those patients with unresectable hepatocellular carcinoma (HCC). The radiomic model performed similarly to the clinical-radiomics model, suggesting that clinical factors failed to produce incremental value to the radiomics model. It is encouraging to see an article focused on developing treatment response prediction models for targeted patients with HCC. This study offers valuable insights for clinicians on how to use radiomics for making clinical decisions and tailoring individualized treatments for patients with unresectable HCC. While the study’s results are promising, several issues warrant further clarification.

First, while manual segmentation of HCC lesions may achieve the highest accuracy, it is time-consuming and can lead to significant interobserver and intraobserver variability, even for the same lesion. Reproducibility and generalizability issues are major challenges in radiomics research.2 The intraclass correlation between various manual segmentations should be evaluated during feature selection to filter out unstable radiological features affected by manual delineation. High-performing fully automated segmentation models are also recommended to enhance reproducibility and clinical application. We anticipate that advanced deep learning model-based segmentation techniques, such as the large language model MedSAM, will be used for automated liver tumor segmentation.3

Second, we found that when comparing the radiomics model to the clinical-radiomics model using the DeLong test in both the training and validation cohorts, there was no significant difference in the discrimination power. Investigating the correlation between radiomic features or radiomic model scores and variables used in the clinical model (such as hepatitis B surface antigen (HBsAg) positivity and tumor burden score grade) may provide insights into this finding. Moreover, since the study is limited to a single-center investigation, validating the model’s robustness with independent cohorts is crucial. The performance of the models showed that the sensitivity and specificity of the radiomic model were not inconsistent in the training and validation cohorts (training cohort: sensitivity=68.6%, specificity=98.0%; validation cohort: sensitivity=95.2%, specificity=72.4%). While the authors’ use of the clinical impact curve to assess the model’s potential clinical application is commendable, further discussion could enhance understanding. Specifically, if the identified patient subgroups were used in clinical practice, selecting appropriate risk thresholds would need to account for clinical feasibility, healthcare resource constraints, and patient acceptance and preference. Additionally, cost-effectiveness analyses for these refined populations would be valuable.

Third, there is an error in the discussion regarding the use of radiomics and clinical-radiomics models for predicting overall survival (OS) and progression-free survival (PFS) risk stratification. It is important to ensure that the same cut-offs are used for risk stratification in both the training and validation data sets. This is crucial because, when applying the model, it evaluates individual patients rather than a set of patient data. Additionally, including the radiomics model score in both univariate and multivariate analyses would be valuable to determine whether it is an independent prognostic factor for OS and PFS.

Fourth, while radiomic models often depend on numerous computational features, the biological meaning of the features can be unclear, which may impede model interpretability. Connecting quantitative imaging features with biological processes through interdisciplinary technologies can bridge the gap between proof-of-concept and clinical applicability. As mentioned in the work by Hua et al,1 future studies will investigate the prognostic value and underlying mechanisms of conversion therapy by integrating radiomics with tissue section analysis. However, interpretability becomes even more critical when radiomics-enabled imaging biomarkers are used to optimize therapy, as therapeutic decisions driven by biomarkers must be based on pathophysiological interpretations.4 It is expected that future research will focus on integrating biological significance into radiomics, including the analysis of biological pathways and genomic correlations. The biological interpretability of radiomics models enhances their utility and integration into clinical practice.

In conclusion, given the substantial cost associated with LPI therapy, the clinical validation of accurate biomarkers is crucial for selecting patients who will benefit most from the treatment, aiming to maximize efficacy and minimize adverse effects.5 The manuscript presents a promising non-invasive radiomics model for identifying potential candidates for LPI therapy and guiding clinical decision-making. Implementing the proposed improvements would further refine the authors’ research, thereby enhancing its overall clinical impact. We eagerly anticipate more insightful work from these authors in the future.