Figure 1

Conceptual framework for decoding intratumoral heterogeneity of hepatocellular carcinoma (HCC) using multimodal ultrasound. HCC is depicted as a spatially heterogeneous biological mosaic composed of distinct tumor habitats. Four complementary acoustic dimensions are interrogated by multimodal ultrasound: tumor perfusion using contrast-enhanced ultrasound (CEUS), biomechanical properties using shear-wave elastography (SWE), tissue microstructure using quantitative ultrasound and radiofrequency (QUS/RF) analysis, and vascular architecture using superb microvascular imaging (SMI) or ultrasound localization microscopy (ULM). Integration of these complementary imaging biomarkers through acoustic radiomics and artificial intelligence enables the generation of a multimodal acoustic heterogeneity signature, with potential applications in noninvasive prediction of microvascular invasion (MVI), histological grade, recurrence risk, treatment response, and longitudinal monitoring for precision management of HCC."

Figure 2

Whole-tumor versus habitat analysis for imaging intratumoral heterogeneity. (A) Conventional whole-tumor radiomics treats the lesion as a single region of interest, averaging quantitative imaging features across the entire tumor and thereby potentially obscuring spatial biological heterogeneity; (B) Habitat analysis preserves spatial information by clustering multiparametric imaging data into distinct imaging habitats that may correspond to biologically different tumor regions, including hyperperfused, transitional, hypoperfused, and necrotic compartments; (C) Conceptual extension to multimodal ultrasound, in which contrast-enhanced ultrasound (CEUS), shear-wave elastography (SWE), quantitative ultrasound/radiofrequency (QUS/RF) analysis, and superb microvascular imaging or ultrasound localization microscopy (SMI/ULM) are spatially integrated to generate region-specific acoustic habitats and a comprehensive multimodal acoustic heterogeneity signature for biological phenotyping of hepatocellular carcinoma. Schematic illustration."

Figure 3

Representative contrast-enhanced ultrasound (CEUS) findings in hepatocellular carcinoma (HCC). A 72-year-old man with HCC confirmed by surgical histopathology (Edmondson–Steiner Grade II). (A) Grayscale ultrasound with color Doppler imaging demonstrates a heterogeneous hyperechoic liver lesion with peripheral vascular signals; (B, C) Following intravenous administration of 2.0 mL SonoVue®, the lesion exhibits heterogeneous arterial-phase hyperenhancement with a characteristic mosaic enhancement pattern; (D) Time-intensity curve analysis demonstrates peak enhancement at 23 s. Washout begins at approximately 45 s, followed by persistent hypoenhancement during the portal venous and late phases, consistent with the typical CEUS enhancement pattern of HCC. This case illustrates the marked intratumoral perfusion heterogeneity that can be visualized by CEUS and forms the basis for quantitative perfusion analysis and acoustic radiomics."

Table 1

Representative reported performance of imaging biomarkers for preoperative prediction of microvascular invasion in hepatocellular carcinoma (non-head-to-head comparison)"

Figure 4

Representative reported diagnostic performance of imaging biomarkers for the preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) (non-head-to-head comparison). Dots represent reported areas under the receiver operating characteristic curve (AUROCs) or pooled estimates from published studies, whereas horizontal bars indicate 95% confidence intervals or the range between internal and external validation when available. The dashed vertical line denotes an AUROC of 0.80, a commonly accepted threshold for good diagnostic discrimination. Although several imaging approaches-including CT/MRI radiomics, contrast-enhanced ultrasound (CEUS), ultrasound radiomics, deep learning, and radiofrequency (RF)-based analysis-have demonstrated promising diagnostic performance, the decline observed between internal and external validation of ultrasound deep-learning models highlights the persistent challenge of model generalizability. Due to differences in study design, patient populations, reference standards, and validation strategies, the parallel display in the figure is only used to illustrate the research background, and the ranking is based solely on reported values and does not represent the level of evidence."