Hepatocellular carcinoma (HCC) is one of the three leading cancers with the lowest survival rates worldwide (1). In China, liver cancer is burdened and has an insidious onset, and most HCC patients are in advanced stages at presentation, particularly those with cirrhosis or severe liver fibrosis, often losing the opportunity for surgical resection (2). Tyrosine kinase inhibitors (TKIs) and other systemic treatments were made the preferred choice for patients with advanced hepatocellular carcinoma (aHCC) (3, 4). However, the molecular biological and genetic changes during the division of cancer cells endowed HCC with heterogeneous characteristics (5, 6), which affected the therapeutic effects and prognosis of the patients (7).

Epidermal growth factor receptor (EGFR) is located on the cell membrane surface and is a receptor for cell proliferation and signal transduction, and its expression status is related to tumor progression and prognosis (8, 9). EGFR high expression (EGFR^high) status activates more downstream signaling pathways and promotes proliferation, metastasis and invasiveness of tumor cells, resulting in poor tumor prognosis (10, 11). EGFR is highly expressed in 40–70% of HCC patients, and it has been shown that HCC patients with EGFR^high have a poor prognosis and have a shorter survival time than those with low EGFR expression (EGFR^low) (12, 13). A recent study published in Nature found that HCC patients with EGFR^high were more likely to develop resistance to TKIs, particularly lenvatinib (12). Only in HCC patients with EGFR^high, lenvatinib induces the feedback activation of EGFR and its downstream PAK2-ERK5 signaling pathway by inhibiting FGFR and downstream ERK1/2 (14), and simultaneously activates the downstream signaling pathway MEK1/2-ERK1/2, which is common with FGFR, resulting in strong proliferation ability of HCC cells while lenvatinib was administered. EGFR inhibitors effectively blocked feedback activation, and combined with Lenvatinib, produced synergistic antitumor effects, indicating that HCC patients with EGFR^high could benefit from this combination (12, 15). Thus, prediction of EGFR expression status not only allows assessment of HCC prognosis, but also enables precise treatment strategies for risk stratification of patients.

EGFR expression requires immunohistochemical detection by surgical resection specimens or biopsies. However, such invasive and less reproducible modality is not suitable for patients with aHCC. Radiomics is a specific algorithm that performs feature extraction and deep mining of standard medical images not only quantifies image features, but also analyzes the molecular phenotype of tumor cells to explore tumor heterogeneity in a non-invasive and reproducible manner (16, 17). Previous researches have reported that computed tomography (CT), magnetic resonance imaging (MRI) and Ultrasound (US) based on radiomics features have the ability to noninvasively characterize biomarkers such as cytokeratin 19 (CK19), vascular endothelial growth factor receptor (VEGFR), and P53 and have achieved promising predictive results (18–20). Up to now, there are few reports on the use of radiomics features to predict EGFR expression in patients with HCC. Since ultrasound is non-invasive, non-radiative, highly repeatable, and reasonably priced, it is one of the most often used techniques for liver testing (20). Thus, the current study is intended to investigate the value of ultrasomics features based on gray-scale ultrasound images for noninvasive prediction of EGFR expression status in patients with HCC, thus providing more objective evidence for precise treatment of aHCC.

HCC is the most common type of liver cancer, with poor prognosis and 5-year survival rate of 18%, and is one of several malignancies with high fatality rate worldwide (3). Multiple molecular biomarkers such as EGFR, ki-67, VEGF, P53 have been identified as the main factors involved in HCC progression and affecting prognosis (24, 26–28), of which EGFR is often highly expressed in HCC and is involved in proliferation, invasion and metastasis of tumor cells, resulting in poor prognosis of HCC (29–31). EGFR was found highly expressed in 50.5% of HCC patients in the present study, which is close to previous findings (8, 10).

Recently, some studies have found that EGFR^high in liver cancer cells is associated with resistance to targeted agents such as Lenvatinib (32, 33). Lenvatinib has been approved by Food and Drug Administration (FDA) as first-line treatment for aHCC. However, the objective response rate was only 24.1%, indicating lenvatinib needs to be combined with other drugs to improve its clinical benefit (34, 35). A study conducted by Jin et al. (12) found that lenvatinib, in HCC patients with high EGFR expression, induced feedback activation of EGFR and its downstream signaling pathways by inhibiting FGFR, leading to HCC cells still having strong proliferative capabilities. EGFR inhibitors could block the feedback-activated signaling pathways, enhancing the antitumor effect. A clinical trial initiated by Renji Hospital (NCT04642547) recruited aHCC patients with high EGFR expression, using a combination treatment of lenvatinib and EGFR inhibitors, and the clinical response rate reached 50% (12, 15, 36). Therefore, HCC patients with noninvasive identification EGFR^high are important conditions for therapeutic management.

In the present study, we compared the predictive performance of five machine learning algorithms constructed clinical models, ultrasomics models, and combined models for noninvasive prediction of EGFR expression in HCC patients. The findings demonstrated that the five machine learning algorithms’ ultrasomics features could successfully differentiate between the EGFR expression status of HCC patients in the training and test datasets (p < 0.05), and that the ultrasomics model and the combined model constructed with an RF classifier outperformed the others in terms of predictive performance (Figures 4E,F; Table 3).

The ultrasomics models developed according to five machine learning algorithms showed good predictive ability, with the RF classifier having the best predictive performance, and the AUC of the training and test datasets increased from 0.846 (95%CI, 0.776–0.902) and 0.713 (95%CI, 0.582–0.823) to 0.929 (95%CI, 0.874–0.966) and 0.807 (95%CI, 0.684–0.897), respectively, for the clinical model. The improvement in predictive performance is because ultrasomics can extract more features from images that are associated with tumor heterogeneity and assess them quantitatively (17, 37, 38). Wu et al. developed an radiomics model based on energy-enhanced CT to predict EGFR expression status in peripheral lung cancer (39). Features such as the arterial phase Laplace of Gaussian Filter Glszm Small Area Low Gray Level emphasis and wavelet HHL gray level co-occurrence matrix (GLCM) MCC, and the venous phase wavelet LHL first-order root mean square were extracted. A multiphasic model established based on the features from both phases was found to have good predictive performance (AUC 0.950). The results show that imaging features, especially higher-order features, can better predict the expression of EGFR. In this study, 7 of the 11 best features are high-order features obtained by wavelet filtering, which indicates that higher-order features can obtain more EGFR-related features, which is the same as previous research results. The seven wavelet transform features were primarily derived from gray-level size-zone matrix (GLSZM) features and first-order features. GLSZM was used to describe the spatial distribution of gray level values and the information about the size of regions in the image, while First-order features mainly described the symmetry, uniformity, and distribution changes of image voxels’ intensity (38, 40). Among the 11 features, the wavelet HLH first-order Minimum, wavelet LHL first-order Median, square gray-level run length matrix (GLRLM) RunEntropy, and wavelet HHL GLSZM Size Zone NonUniformity features were found to have the highest coefficients (Figure 6). The square feature involves squaring each pixel value in the image to enhance the contrast of gray level values (41). These results indicate that wavelet transform features could increase the predictive value of radiomic features, being more sensitive to the identification of tumor heterogeneity and can be used to predict the EGFR expression status (42). When clinical baseline characteristics were incorporated into the ultrasomics model, the RF machine learning combined model demonstrated better predictive performance, with AUCs of 0.937 (95%CI, 0.884–0.971) for the training dataset and 0.822 (95%CI, 0.702–0.909) for the test dataset, showing a slight improvement over the ultrasomics model alone. Therefore, the ultrasomics model and combined model established by the RF classifier can better predict the EGFR expression status in HCC patients.

Qualitative characteristics of lesion margin on ultrasound images analyzed by multivariate logistic regression may serve as independent predictors of EGFR expression status. Unclear lesion boundary is a risk factor, which may be related to EGFR^high more aggressive and proliferative ability leading to changes in lesion morphology (40, 43). In univariate analysis, there was a significant difference between groups in serum AFP and NLR (p < 0.05), and correlation analysis showed that the correlation coefficients (r) between AFP and NLR and EGFR expression status were 0.146 and 0.154, respectively (p < 0.05) (Supplementary Figure S3), indicating that serum AFP and NLR had a low correlation with EGFR expression status. Fan et al. predicted VEGF expression in HCC patients based on MRI imageomics profiles, and multivariate logistic regression analysis showed that AFP, NLR, and irregular lesion boundaries were independent predictors, and AUC of the clinical model and imageomics model in the training and test dataset were 0.709, 0.725; 0.892, 0.800, respectively (44). The AUC of the RF machine learning clinical model in this study was 0.846 (95%CI, 0.776–0.902) and 0.713 (95%CI, 0.582–0.823) in the training and test dataset, respectively, and compared to the ultrasomics model, the prediction accuracy was lower (AUC 0.929, 0.807). These data suggested that clinical baseline characteristics have limited predictive power for molecular tumor phenotypes.

Due to the limitations of EGFR immunohistochemical testing in the patients, in this present study, there was just a single center involved and lacked external validation to enhance the generalizability of the predictive model. Another limitation of the study was that manual segmentation was time-consuming and inefficient. Thus, in the future, there is a need for convenient, efficient, and repeatable automatic segmentation software that must be clinically validated. Third, the baseline images in this experiment were taken from different ultrasound instruments. Although the image is preprocessed before feature extraction, there may be confounding factors that affect the results. Finally, the present study only analyzed gray-scale ultrasound and did not assess in conjunction with contrast-enhanced ultrasound, elastography, and other imaging modalities. In our future work, multimodal imaging radiomics will be explored for the EGFR expression levels in HCC patients.

In conclusion, the construction of ultrasomics based on gray-scale ultrasound images by five machine learning algorithms can be used as noninvasive and effective diagnostic tools to predict EGFR expression status in HCC patients. Furthermore, the ultrasomics model and combined model established by RF classifier have the best predictive performance. The present study will provide a new noninvasive method for noninvasive prediction and precise treatment of EGFR expression status in patients with aHCC.