Value of radiomics models in precision diagnosis of dual-phenotype hepatocellular carcinoma and intrahepatic cholangiocarcinoma

Abstract

In the clinical diagnosis and treatment of liver tumors, the differential diagnosis between dual-phenotype hepatocellular carcinoma and intrahepatic cholangiocarcinoma has long been a challenging problem for clinicians. These two types of tumors not only exhibit overlapping pathological features but also have significantly different treatment strategies and prognoses. Misdiagnosis can directly affect patients' quality of life. Recently, a study published by Zhang et al has brought groundbreaking insights into this dilemma. Radiomics technology has demonstrated remarkable value in the differential diagnosis of these two diseases, opening up a new path for the precise diagnosis and treatment of liver tumors.

Key Words: Dual-phenotype hepatocellular carcinoma; Intrahepatic cholangiocarcinoma; Radiomics; Differential diagnosis; Machine learning

Core Tip: A recent study published by Zhang et al focuses on exploring the practical value of radiomics in ahcieving the differential diagnosis between dual-phenotype hepatocellular carcinoma and intrahepatic cholangiocarcinoma. Its constructed combined model, with an area under the receiver operating characteristic curve of 0.892, can effectively boost diagnostic accuracy and provide significant assistance to less-experienced radiologists in clinical practice. However, the study still faces limitations of single-center data and small sample size, and future research thus needs to carry out multi-center studies and promote multi-omics fusion to address these issues.

TO THE EDITOR

Dual-phenotype hepatocellular carcinoma (DPHCC) and intrahepatic cholangiocarcinoma (ICC) are two clinically important subtypes of primary liver cancer, but their differential diagnosis has long plagued clinicians due to overlapping pathological and imaging features[1,2]. DPHCC is a rare special subtype of hepatocellular carcinoma, characterized by simultaneous differentiation of hepatocyte and cholangiocyte lineages; it has stronger biological aggressiveness, higher rates of microvascular invasion, and poorer sensitivity to radiotherapy and chemotherapy than conventional HCC, leading to significantly worse prognosis[3]. In contrast, ICC originates from intrahepatic bile duct epithelial cells, spreads mainly through lymph node metastasis, and its standard treatment relies on surgical resection combined with postoperative adjuvant chemotherapy[4]. Misdiagnosis between DPHCC and ICC directly leads to inappropriate treatment selection: For example, treating DPHCC with ICC-oriented adjuvant chemotherapy may result in ineffective therapy and missed opportunities for targeted intervention, while misclassifying ICC as DPHCC may lead to unnecessary aggressive surgical approaches and increased postoperative complications[2,5]. Traditional imaging diagnosis (such as computed tomography and magnetic resonance imaging) relies on qualitative description of tumor morphology, vascular enhancement patterns, and peritumoral changes. In this context, radiomics, which extracts quantitative features from medical images and constructs diagnostic models via machine learning, has emerged as a promising tool to address this diagnostic dilemma[6].

Study design and clinical significance

A study recently published by Zhang et al has brought groundbreaking insights into this dilemma. This study constructed a clinical sign model, a radiomics model, and a combined model, and conducted an in-depth analysis of the imaging data from 53 DPHCC patients and 124 ICC patients. The results showed that the area under the curve of the combined model in the test set reached 0.892, significantly superior to that of the single clinical model. More notably, with the assistance of the model, the diagnostic area under the curve of less experienced radiologists (with 5 years of experience) increased from 0.698 to 0.946, confirming the unique role of radiomics in bridging the gap in diagnostic experience. This finding not only validates the clinical applicability of radiomics but also reveals its great potential as a “digital biopsy” tool. The core value of radiomics lies in converting texture features in medical images that are difficult to capture with the naked eye into quantitative data and uncovering the essence of diseases through machine learning algorithms[7]. Compared with traditional imaging diagnosis, this study adopted whole-tumor slice-by-slice delineation of regions of interest, avoiding the limitations of single-slice analysis and comprehensively capturing tumor heterogeneity, which reflects the imaging manifestations of molecular-level differences between DPHCC and ICC. The 28 radiomic features screened in the study may represent the “imaging fingerprints” of the biological behaviors of these two tumors, such as cell proliferation and angiogenesis.

Zhang et al’s work has some notable innovations compared with earlier studies on DPHCC-ICC differential diagnosis[1]: (1) Whole-tumor regions of interest delineation: Unlike Gu et al[8] (who focused on tumor parenchyma only), Zhang et al[1] adopted slice-by-slice delineation of the entire tumor, including the tumor margin and peritumoral tissues. This approach comprehensively captures tumor heterogeneity, which better reflects the molecular-level differences between DPHCC (with hepatocyte-cholangiocyte dual differentiation) and ICC (with bile duct epithelial differentiation); and (2) feature interpretability clues: The 28 screened radiomic features included 12 texture features (e.g., gray-level co-occurrence matrix entropy) and 8 shape features (e.g., tumor sphericity), which may correspond to biological behaviors such as cell proliferation (reflected by texture irregularity) and angiogenesis (reflected by enhancement heterogeneity) in DPHCC and ICC. This provides a basis for linking radiomic features to pathological mechanisms, whereas Gu et al’s study (focused on magnetic resonance imaging qualitative features) lacked such quantitative-pathological correlation analysis[8].

From a practical clinical perspective, radiomics models are particularly valuable for physicians at different experience levels. For experienced doctors, the model can serve as a reference to help reduce subjective bias in diagnosis; for young, less experienced doctors, radiomics models can help them quickly improve their diagnostic capabilities. Currently, the diagnosis and treatment of liver tumors are moving towards precision and personalization[9]. As a special subtype of HCC, DPHCC has characteristics of both hepatocyte and cholangiocyte differentiation, with stronger microvascular invasion ability and higher resistance to radiotherapy and chemotherapy[7]. In contrast, ICC primarily spreads through lymph node metastasis, and adjuvant chemotherapy after surgery is the standard treatment[10]. Precise differentiation between the two directly determines the choice of surgical approach and adjuvant treatment strategies. The introduction of radiomics models not only improves the accuracy of preoperative diagnosis but also provides objective quantitative evidence for clinical decision-making, promoting the shift from “experience-based medicine” to “data-driven medicine”.

Existing challenges and future directions

However, Zhang et al’s study also highlights common challenges in the clinical translation of radiomics. The single-center retrospective design may introduce selection bias, the limited sample size may affect the generalization ability of the model, and the integration of imaging data with multi-omics information remains to be further explored. In the future, multi-center large-sample studies, cross-modal data fusion (such as integration with genomics and proteomics), and the development of interpretable artificial intelligence algorithms will be crucial for overcoming these bottlenecks. Model interpretability can be improved by integrating SHapley Additive exPlanations into radiomics models, an approach that quantifies the contribution of each feature to the diagnostic result[11]. As medical imaging evolves from “qualitative description” to “quantitative analysis” and artificial intelligence transitions from the laboratory to the clinical frontline, radiomics is reshaping the diagnostic paradigm for liver tumors. This study not only provides a practical tool for the differential diagnosis of DPHCC and ICC but also heralds the advent of a new era where "imaging serves as a biomarker". With the continuous iteration of technology and the innovation of concepts, radiomics is bound to play an even more significant role in tumor early screening, treatment efficacy assessment, prognosis prediction, and other fields.