Advancing precision in hepatocellular carcinoma prognostication: The promise of biparametric magnetic resonance imaging-based multimodal models

Abstract

Zuo and Liu investigated the value of a novel noninvasive approach integrating biparametric magnetic resonance imaging, radiomics, deep transfer learning, and clinical factors in predicting Ki-67 risk stratification and recurrence-free survival (RFS) in hepatocellular carcinoma (HCC). The study included 198 HCC patients and utilized histopathological Ki-67 expression as the reference standard for risk stratification. The integrated multimodal model combining radiomic features, deep transfer learning signatures, and clinical factors (nonsmooth tumor margin and absence of an enhanced capsule), achieved an area under the curve of 0.92 in the training and validation cohorts for predicting high Ki-67 risk, with a sensitivity and specificity of 0.88 and 0.85, respectively. Furthermore, the model effectively stratified RFS, with median RFS of 33.53 months in the high-risk group vs 66.74 months in the low-risk group, consistent with histopathological findings that directly refer to Ki-67 stratification. The findings highlight the potential of biparametric magnetic resonance imaging-based multimodal models in noninvasive HCC prognostication, though external validation in larger cohorts is warranted. The demand for precise, noninvasive preoperative assessment tools in HCC management remains high in clinical practice.

Key Words: Hepatocellular carcinoma; Ki-67; Biparametric magnetic resonance imaging; Radiomics; Deep transfer learning; Recurrence-free survival

Core Tip: Zuo and Liu examined the prognostic performance of a biparametric magnetic resonance imaging-based multimodal model in hepatocellular carcinoma. By integrating radiomics, deep transfer learning - a technique that captures subtle imaging patterns imperceptible to the human eye through convolutional neural networks, and clinical factors, this integrated multimodal model effectively predicts Ki-67 risk stratification and recurrence-free survival, offering a noninvasive alternative to invasive histopathological analysis. Given the retrospective nature of the cohort, further validation in multicenter, prospective studies is essential. This work advances precision prognostication in hepatocellular carcinoma, addressing the critical need for preoperative risk stratification tools.

TO THE EDITOR

In the World Journal of Hepatology, Zuo and Liu[1] introduced a novel multimodal approach integrating biparametric magnetic resonance imaging (bpMRI) with radiomics, deep transfer learning (DTL), a technique that leverages pre-trained convolutional neural networks to capture subtle, high-dimensional imaging patterns imperceptible to human observers, and clinical factors to predict Ki-67 risk stratification and recurrence-free survival (RFS) in hepatocellular carcinoma (HCC). HCC remains a leading cause of cancer-related mortality globally, characterized by aggressive progression, high recurrence rates, and limited long-term survival despite advances in surgical and locoregional therapies. Preoperative identification of tumors with high proliferative potential is critical for optimizing treatment strategies, as it directly influences decisions on neoadjuvant therapy, surgical extent, and surveillance intensity[2].

The Ki-67 index, a well-validated marker of cellular proliferation, is widely recognized as a key prognosticator in HCC: High Ki-67 expression (> 20%) correlates with increased recurrence risk and shortened survival[3]. However, its assessment relies on invasive histopathological analysis of tumor tissue, which is unavailable preoperatively and limits its utility in treatment planning. This gap underscores the need for noninvasive tools to predict Ki-67 status and associated clinical outcomes.

Several guidelines recommend imaging-based assessment of HCC aggressiveness, including multisequence magnetic resonance imaging (MRI), contrast-enhanced computed tomography, and ultrasound[4]. Among these, MRI offers superior soft-tissue resolution, with radiomic features derived from MRI sequences increasingly used to quantify tumor heterogeneity and predict biological behavior. Traditional radiomic models, however, often depend on complex, time-consuming multisequence protocols, limiting their accessibility in resource-constrained settings.

Zuo and Liu[1] addressed this limitation by focusing on bpMRI, which includes T2-weighted imaging and arterial-phase imaging, two core sequences that balance efficiency and diagnostic value. By streamlining the protocol, bpMRI reduces scan time and healthcare costs without sacrificing key tumor characteristics. The authors further enhanced predictive power by integrating radiomics and DTL: Radiomics extracts quantifiable features (e.g., texture and shape) reflecting tumor heterogeneity, while DTL (via convolutional neural networks such as ResNet101) captures complex patterns imperceptible to human observers.

A rigorous feature selection process, including Z-score normalization, t-tests, Spearman correlation, and least absolute shrinkage and selection operator - a statistical method for feature selection and regularization, ensured the retention of robust signatures, mitigating overfitting risks. The resulting radiomic model [area under the curve (AUC) = 0.81] and DTL model (AUC = 0.87) outperformed the clinical model (AUC = 0.77) in predicting high Ki-67 risk, highlighting the value of quantitative imaging over qualitative radiological assessment alone. The integrated nomogram, which combined these imaging signatures with clinical imaging features (nonsmooth tumor margin and absence of an enhanced capsule), achieved an AUC of 0.92 in the training and validation cohorts, demonstrating superior discriminative ability.

Notably, the nomogram’s prognostic utility was validated by its ability to stratify RFS: Patients predicted to have high Ki-67 risk had a median RFS of 33.53 months, significantly shorter than the 66.74 months observed in the low-risk group[1]. This aligns with histopathological findings specifically based on Ki-67 stratification, where high Ki-67 expression correlated with shorter RFS (33.00 months vs 66.73 months), confirming the model’s clinical relevance. The strengths of this study lie in its multimodal design and focus on practicality. By leveraging bpMRI, the model reduces barriers to implementation, making it feasible for widespread use. The integration of radiomics and DTL capitalizes on the strengths of both approaches: Radiomics provides interpretable quantitative features, while DTL captures high-dimensional patterns, together enhancing predictive accuracy[5].

Despite these advances, several limitations merit attention. The retrospective design, while including a relatively large cohort (198 patients), lacks external validation - a critical step for translating such models into clinical practice. The discrepancy in performance between the training and validation cohorts (e.g., radiomic model AUC = 0.81 vs 0.65) raises concerns about generalizability, potentially due to overfitting or variability in imaging protocols. Additionally, bpMRI, while efficient, may overlook informative features in portal venous or delayed phases, which are linked to microvascular invasion and tumor differentiation[6].

Another area for refinement is model interpretability. While gradient-weighted class activation mapping heatmaps offer insights into DTL focus areas, clarifying how specific radiomic or DTL features correlate with Ki-67 expression (e.g., texture heterogeneity reflects underlying proliferative activity) would strengthen biological plausibility. Techniques like SHapley Additive exPlanations could help demystify these relationships, fostering clinician trust[7]. Larger, multicenter prospective studies are necessary to validate the nomogram in diverse patient populations. Incorporating additional sequences (e.g., diffusion-weighted imaging) or modalities (e.g., contrast-enhanced ultrasound) may further enhance performance. Embedding the model into user-friendly clinical software would facilitate real-time preoperative risk assessment, enabling personalized treatment decisions.

CONCLUSION

Noninvasive prediction of Ki-67 risk stratification and RFS in HCC is critical for optimizing treatment strategies. Zuo and Liu’s multimodal model[1], based on bpMRI, radiomics, DTL, and clinical factors, demonstrates potential in addressing this need. Further validation in larger cohorts with clinical outcomes is essential to confirm its utility[1]. The persistent clinical demand for precise, noninvasive prognostic tools in HCC management remains high, and this work represents a significant step toward fulfilling that critical clinical need.