Preoperative Prediction of Macrotrabecular-Massive Hepatocellular Carcinoma Using Machine Learning-Based Ultrasomics

Introduction

Hepatocellular carcinoma (HCC) is the most prevalent form of primary liver cancer, representing about 75–85% of cases and ranking as the third leading cause of cancer deaths globally.¹ Despite significant advances in HCC treatment, the overall prognosis of HCC remains poor,² largely due to the potential genetic alterations, molecular characteristics, and histological patterns.³ According to the fifth edition of the World Health Organization classification of tumors in 2019, approximately 35% of HCCs are special pathological subtypes, including steatohepatitic, clear cell type, MTM, scirrhous, chromophobe, fibrolamellar, neutrophil-rich, and lymphocyte-rich subtypes.⁴ Recent reports have shown that MTM-HCC is an invasive HCC subtype^5–7 closely associated with TP53 gene mutations, fibroblast growth factor 19 (FGF19) amplification, and the ataxia -telangiectasia mutated (ATM) protein.⁸ Meanwhile, this subtype is an independent predictor of early and overall recurrence in patients who have undergone surgical resection⁹ and is linked to poor clinical prognosis.^5,10 Liu et al reported high PD-L1 expression in the MTM-HCC subtype, suggesting that patients with MTM-HCC may benefit from tumor immunotherapy.¹¹ Accurate preoperative diagnosis of MTM-HCC is of great significance to the choice of clinical treatment methods and the improvement of prognosis. Although pathological examination can accurately diagnose MTM-HCC before operation, there are risks such as invasiveness, sampling error, and possible complications. Therefore, a non-invasive, simple and reproducible examination method is urgently needed to help clinical preoperative diagnosis of MTM-HCC.

Radiomics is an emerging technology that enables the high-throughput extraction of medical image features, followed by the selection of key features using dimensionality reduction, to construct models for evaluating information related to diagnosis, treatment response and prognosis of oncological diseases.^12–17 For example, Hawkins et al found that RF classification-based CT imaging can be used to predict the occurrence of lung cancer nodules.¹² The peritumor vascular and intratumoral radiomics model developed by Xie et al provided a non-invasive tool to predict pathologic complete responses in triple-negative breast cancer patients receiving neoadjuvant chemoradiotherapy.¹³ Regarding the application of radiomics to predict MTM-HCC, there have been preliminary studies in the direction of CT and, MRI,^18–20 and achieved good prediction results. For example, Feng et al collected contrast-enhanced CT image data and developed an SVM-based radiomics model based on the training set. The authors found that the AUCs of the model in the training set, internal test set, and external test set were 0.84, 0.80, and 0.74, respectively.¹⁸

However, compared with CT and MRI, ultrasound is simple, radiation-free, portable and lower cost. In recent years, AI-driven ultrasound technology has been further developed and applied in clinical practice to reduce subjectivity and improve the efficiency of ultrasound diagnosis,²¹ and many studies have confirmed its value in liver diseases. Peng et al used dimensionality reduction techniques and machine learning methods to develop ultrasonomics models that can be used to predict HCC, intrahepatic cholangiocarcinoma and combined hepatocellular-cholangiocarcinoma in preoperative noninvasive differentiation.²² Ren et al constructed an SVM-based algorithm to construct an ultrasomics model that can be used to help in preoperative noninvasive prediction of HCC pathological grade.²³ Dong et al²⁴ published a study that the SVM algorithm based on grayscale ultrasound images has the potential to help predict preoperative microvascular invasion status in HCC. However, the accuracy of ultrasomics in the preoperative prediction and differential diagnosis of MTM-HCC is currently unclear. This study developed different ultrasomics models using ML to predict MTM-HCC before surgery, offering a new approach for non-invasive, early diagnosis and treatment selection for the disease.

Material and Methods

Subjects

This retrospective study was approved by the Medical Ethics Committee of Henan Provincial People’s Hospital (2021 Ethics Review No. 01) and exempted the patient from informed consent.

From January 2021 to December 2023, Patients with pathologically confirmed HCC were enrolled consecutively. The detailed eligibility criteria were as follows:

Inclusion criteria: (1) pathologically confirmed hepatocellular carcinoma, with complete pathological data; (2) abdominal ultrasound performed two weeks before surgery, with available ultrasound images; (3) underwent radical resection for liver cancer. Exclusion criteria: (1) prior radiotherapy, chemotherapy, radiofrequency ablation or other antitumor treatments; (2) other comorbid malignancies; (3) Patients with unclear ultrasound images and incomplete lesion display; (4) lack of complete clinical pathological data. Finally, 211 patients were enrolled in the study. The case screening process is shown in Figure 1.

Figure 1 Screening and enrollment of cases according to established eligibility criteria.

Clinical Data

Clinical data that were collected included age, sex, history of hepatitis, cirrhosis, splenomegaly, ascites, portal hypertension, tumor biomarker (alpha-fetoprotein, AFP), aspartate aminotransferase (AST), alanine aminotransferase (ALT), albumin (ALB), globulin (GLO), albumin-globulin ratio (A/G), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), alkaline phosphatase (ALP), creatinine (CREA), activated partial thromboplastin time (APTT), prothrombin activity (PTA), prothrombin time (PT), fibrinogen (FIB), and thrombin time (TT).

Pathological Data

All histological sections were reviewed by a senior pathologist who was unaware of the patients’ clinical and imaging data. MTM-HCC was defined as tumors with a macrotrabecular growth pattern (trabeculae more than six cells thick) observed in more than 50% of the tumor tissue according to histologic diagnostic criteria.²⁵

Image Acquisition and Segmentation

Patients were required to fast for more than 8 hours before examination. Each patient was examined by ultrasonographers with more than 10 years of experience, and ultrasound images were saved in Digital Imaging and Communications in Medicine (DICOM) format. To reduce sample selection bias, the largest lesion was selected as the target lesion for patients with multiple lesions. Ultrasound images showing the largest diameter of the target lesion were analyzed. The model of ultrasound instrument used in the examination: GE Logiq E20, GE Vivid E9, Philips EPIQ 7, HIVISION Ascendus, etc. Choose convex array abdominal probe with frequency of 1–5MHz.

To minimize differences between ultrasound equipment and operators and to facilitate subsequent feature extraction and comparison, images were preprocessed by researchers with eight years of experience. During the preprocessing stage, first, image features were standardized using the z-score. Second, the interpolation resampling method was applied to achieve a pixel size of 1mm × 1 mm using B-spline interpolation. The image is grayscale discretized in the histogram, with the bin width fixed at 25, and then feature calculation is performed.

Image segmentation was performed using the open-source software ITKSNAP (http://www.itksnap.org). The region of interest (ROI), was first delineated manually by an ultrasound doctor with 15 years of experience in ultrasound. Thirty randomly selected target lesions were delineated independently by another ultrasound doctor (with 30 years of experience). The intraclass correlation coefficient (ICC) assessed intra-observer consistency, with an ICC≥0.8 indicative of reproducibility.²⁶ A flowchart of the study procedure is shown in Figure 2.

Figure 2 Overall study flow chart. (A) Image acquisition and lesion segmentation (B) Feature extraction and selection (C) Model construction and evaluation.

Feature Extraction and Selection

The original images apply 14 filters to obtain the derived images for each patient. Feature extraction was performed on all original and derived images using Pyradiomics. The extracted features mainly included first-order statistical features, two-dimensional shapes, gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), neighborhood gray-tone difference matrix (NGTDM), and gray level dependence matrix (GLDM). Dimensionality reduction was performed on the features using a combination of variance filtering, mutual information, and the embedding method combined with XGBoost to reduce redundant features and improve modeling efficiency and model accuracy.

Model Construction and Evaluation

Data were divided 7:3 into a training set (n=147) and a test set (n=64) using a stratified sampling method. The screened ultrasound image features were used to construct ultrasomics model. Univariate and multivariate analyses were adopted to screen out statistically significant factors from the clinical data, and constructed clinical model. A combined model was built by integrating both clinical features and ultrasomics features. Five ML algorithms, namely RF, XGBoost, SVM, DT, and LR, were each used to construct the three models for the training set. The performance of the models was evaluated on the test set using AUC, sensitivity, specificity, and accuracy. The working steps are presented in the form of pseudo-code in Figure 3.

Figure 3 The pseudo-code corresponding to the algorithm is shown in Algorithm 1.

Statistical Analysis

Data were analyzed using IBM SPSS Statistics 27.0. Quantitative data are expressed as mean ± standard deviation when normally distributed, and as median (interquartile range) when not normally distributed. These data were compared between groups using the independent t-test or Mann–Whitney U-test. Qualitative data are expressed as numbers(percentages) and compared using the Chi-square test or Fisher’s exact test. P < 0.05 was considered statistically significant.

Results

Clinical and Pathological Features

Of the 211 patients included in this study, 169 were males and 42 were females, with an age of 28–85 years (57.36 ± 10.05 years). Among them, 51 cases (24.2%) were diagnosed with MTM-HCCs and 160 cases (75.8%) with non-MTM-HCCs. There were no noticeable difference between the training set and test set. The baseline clinical and pathological data of the patients are shown in Table 1. Univariate analysis revealed that ALT (P=0.017), AST (P<0.001), AFP (P=0.005), gender (P=0.038), and tumor diameter (P<0.001) were statistically significant between MTM-HCC and non-MTM-HCC groups, as shown in Table 2. The relevant factors were incorporated into a multivariable logistic regression analysis, with results shown in Table 3. The findings indicate that higher levels of AFP (P=0.031), AST (P=0.030), and tumor diameter (P=0.034) are independent risk factors for diagnosing MTM-HCC. The ultrasound images and histological patterns of non-MTM-HCC and MTM-HCC are shown in Figure 4.

Table 1 Clinical and Pathological Data of Patients in Training Set and Test Set

Table 2 Univariate Analysis of Preoperative Data for MTM-HCC and Non-MTM-HCC Patients

Table 3 Multifactorial Analysis of Predicting MTM-HCC

Figure 4 The ultrasound images and histological patterns of non-MTM-HCC and MTM-HCC. (A)–(D): Ultrasound and pathological images of a 65-year-old male patient with MTM-HCC; (A) and (B) Ultrasound images of MTM-HCC patients and ROIs; (C) and (D): Histopathology images of MTM-HCC at low and medium magnification; (E)–(H) Ultrasound and pathological images of a 70-year-old male patient with non-MTM-HCC; (E) and (F) Ultrasound images of non-MTM-HCC patients and ROIs; (G) and (H): Histopathology images of non-MTM-HCC at low and medium magnification.

Feature Extraction and Selection

A total of 1409 features were extracted from the patients’ original and derived ultrasound images. First, 260 features with ICC lower than 0.8 were excluded after intraobserver agreement analysis; variance filtering and mutual information method were used to exclude 16 features with zero variance and 423 features with zero MIC; and 690 features were excluded by embedding method combined with XGBoost. Twenty features with non-zero coefficients were ultimately identified (Supplementary Figures 1 and 2).

Performance of the Clinical, Ultrasomics, and Combined Models

Comparing the five ML algorithms, the AUCs of the clinical, ultrasomics, and combined models in the test set were 0.805, 0.852, 0.856; 0.621, 0.822, 0.841; 0.733, 0.701, 0.852; 0.639, 0.812, 0.825; 0.743, 0.771, 0.860; the accuracies were 0.750, 0.859, 0.859; 0.672, 0.734, 0.813; 0.656, 0.719, 0.734; 0.625, 0.797, 0.766; 0.688, 0.703, 0.750; the sensitivities were 0.733, 0.800, 0.800; 0.733, 0.800, 0.733; 0.667, 0.733, 0.733; 0.667, 0.733, 0.800; 0.867, 0.733, 0.733; and the specificities were 0.755, 0.878, 0.878; 0.653, 0.714, 0.837; 0.653, 0.714, 0.735; 0.612, 0.816, 0.755; 0.633, 0.694, 0.755, respectively. The results showed that the combined models had higher AUCs than the clinical and ultrasomics models built using the five algorithms. Similarly, the ultrasomics model had a higher AUC than the clinical model constructed with RF, XGBoost, DT, and LR. The RF-based combined model and ultrasomics model (AUC of 0.856 and 0.852, respectively) and the LR-based combined model (AUC of 0.860) had higher AUC values. However, The RF-based models exhibited higher accuracy and specificity compared to models constructed using the other four algorithms. Among the three RF-based models, AUC, accuracy, sensitivity, and specificity were higher for the combined model and ultrasound model than for the clinical model. Furthermore, the combined model demonstrated higher AUC and similar accuracy, sensitivity, and specificity compared to the ultrasomics model. Taken together, all five ML-based ultrasomics models were effective in distinguishing MTM-HCC from non-MTM-HCC, and the combined models exhibited superior performance than the clinical and ultrasomics models. Additionally, RF-based models had the best overall performance compared to other ML-based models.

The ROC curves for the three RF-based models are shown in Figure 5, and the results of the models constructed using the five algorithms are shown in Table 4.

Table 4 Prediction Results of Different Models for Training Set and Test Set

Figure 5 ROC curves in RF training set and test set. (A) ROC curves of ultrasomics model based on the ultrasound image features. (B) ROC curves of clinical model based on clinical data. (C) ROC curves of combined model based on clinical data and ultrasomics features.

Discussion

MTM-HCC is a new histological subtype of HCC identified by the WHO in 2019. Compared to non-MTM-HCC, MTM-HCC is associated with a larger tumor volume, elevated preoperative AFP levels, vascular invasion, higher Edmondson-Steiner grade,^18,25,27 and greater invasiveness and metastatic potential. It is an independent predictor for early postoperative recurrence. Studies have shown that MTM-HCC may benefit from neoadjuvant therapy.^11,18,28 However, MTM-HCC can only be diagnosed by postoperative histopathology or invasive preoperative biopsy, limiting the treatment options for patients with this HCC subtype. Therefore, accurate preoperative predictions using non-invasive and effective methods are crucial for guiding optimal personalized treatment plans.

Radiomics is an emerging non-invasive technique that effectively mines image information. In this study, we extracted 1409 ultrasound image features from the original and derived images of each patient, and performed feature dimensionality reduction through a combination of ICC, variance filtering, mutual information, and the embedding method combined with XGBoost. We identified 20 best features and constructed models using RF, XGBoost, SVM, DT, and LR for the preoperative prediction of MTM-HCC. Analysis showed that all five ML-based ultrasomics models effectively predicted MTM-HCC and the combined models had superior predictive performance than the clinical and ultrasomics models. In addition, RF-based models exhibited better overall performance than models constructed using the other four algorithms. The RF-based combined and ultrasomics models had significantly higher AUCs (0.856 and 0.852 vs 0.805), accuracies (0.859 and 0.859 vs 0.750), specificities (0.878 and 0.878 vs 0.755), and sensitivities (0.800 and 0.800 vs 0.733) than the RF-based clinical model. Furthermore, AUC was slightly higher for the combined model compared to the ultrasomics model. Our results indicated that non-invasive ultrasomics is effective for predicting MTM-HCC, highlighting its potential clinical value.

Previous studies have reported that MRI and CT image features are useful for predicting MTM-HCC.^{18–20,29–32} Zhang et al²⁰ found that the AUC of the MRI radiomics model based on the LR algorithm was 0.766 and 0.739, and the specificities were 0.847 and 0.804, respectively. In a multicentre study, Li et al constructed a DL-based dual-energy CT radiomics nomogram based on deep learning (DL) features and manual features. Their model has an AUC of 0.91 in the training set, 0.87 in the internal test set, and 0.89 in the external test set.¹⁹ Our study revealed that the RF-based ultrasomics model had an AUC of 0.938 in the training set and 0.852 in the test set, demonstrating favorable predictive performance. This also indicates that compared with CT and MRI, ultrasomics model features can also provide more information for predicting MTM-HCC. Among the features of constructing ultrasonic models, “Glszone Variance” and “Gldm_Small Dependence High Gray Level Emphasis” had the highest feature coefficients, describing the heterogeneity of texture features in MTM-HCC. It is consistent with some imaging features extracted from previous studies of HCC.²⁰ The abundance of information in grayscale ultrasound images, combined with the radiation-free, simple, and cost-effective characteristics offer significant potential for predicting MTM-HCC.

We found that MTM-HCC patients had higher AST levels than non-MTM-HCC patients (P < 0.05). Shan et al found that higher AST levels were associated with lower survival.³³ The increased release of AST may be attributed to extensive necrosis within the MTM-HCC tumor.

There are also some limitations in this study. The study included only retrospective cases from one hospital, and the sample size of the data was limited, lacking external test sets. As a result, the stability and applicability of the established prediction models could not be fully validated. Therefore, further international multicenter studies are warranted to enhance the effectiveness and generalizability of the model.

Conclusion

In conclusion, the ultrasomics model can distinguish the MTM-HCC subtype through machine learning. The random forest machine learning method has the best comprehensive performance and excellent prediction ability, which will provide a new method for non-invasive identification of MTM-HCC.