B-mode ultrasound and contrast-enhanced ultrasound-based radiomics interpretable analysis for the prediction of macrotrabecular-massive subtype of hepatocellular carcinoma

Dan Lu, Cheng Qin, Lifan Wang, Lingling Li, Yu Li, Liping Sun, Hui Shi, Boyang Zhou, Xin Guan, Miao Yao, Hong Han, Jianhua Zhou, Huixiong Xu, Chong-Ke ZhaoThe Ultrasound Journal 2025 · DOI

This research paper, recently published in The Ultrasound Journal, presents a significant step forward in using artificial intelligence to refine liver cancer diagnosis and prognosis.

TL;DR

Researchers developed an interpretable machine learning model that uses B-mode and contrast-enhanced ultrasound (CEUS) to identify the aggressive "macrotrabecular-massive" (MTM) subtype of liver cancer before surgery. Their AdaBoost-based radiomics model achieved high accuracy (AUC 0.88) and successfully predicted early cancer recurrence, offering a non-invasive path toward personalized surgical planning.

What Problem Does This Paper Address?

Hepatocellular carcinoma (HCC) is not a monolithic disease; it has several subtypes with vastly different outlooks. The macrotrabecular-massive (MTM) subtype is particularly dreaded because it is highly aggressive, frequently invades blood vessels, and has a high rate of recurrence after surgery.

Currently, the "gold standard" for identifying MTM-HCC is a pathology report after the tumor has been surgically removed. While radiologists look for certain signs on CT or MRI scans (like large size or specific enhancement patterns), these interpretations are subjective and prone to "interobserver variability"—meaning two doctors might look at the same scan and disagree. There is a critical need for a non-invasive, objective tool to identify these high-risk patients before they go under the knife.

Methods

The authors turned to radiomics, a field that treats medical images as data. Instead of just looking at a scan, radiomics uses algorithms to extract thousands of quantitative features (texture, shape, intensity) that are invisible to the human eye.

  1. Image Acquisition: They used B-mode ultrasound (standard grayscale) and CEUS, which uses microbubble contrast agents to show blood flow in real-time across three phases: Arterial (AP), Portal Venous (PVP), and Delayed (DP).
  2. Feature Engineering: They extracted 1,070 features per image. After a rigorous four-step selection process (checking for reproducibility and removing redundant data), they narrowed it down to the 47 most "telling" features.
  3. Machine Learning: They tested ten different classifiers. The AdaBoost algorithm emerged as the top performer.
  4. Interpretability: To avoid the "black box" problem common in AI, they used SHAP (SHapley Additive exPlanations). This method quantifies exactly how much each specific feature (e.g., a certain texture in the arterial phase) contributed to the final prediction.

Datasets

The study was multi-institutional, involving three medical centers:

  • Development Cohort: 255 HCC patients from Zhongshan Hospital (Fudan University). This was split 8:2 for training and internal validation.
  • External Test Set: 89 patients from Sun Yat-sen University Cancer Center and Shanghai Tenth People’s Hospital. This external set is crucial for proving the model works on data it hasn't seen before.
  • Demographics: The cohort was predominantly male (~80%) with a mean age of 58, reflecting typical HCC demographics.

Key Insights and Ideas

  • CEUS is the "Secret Sauce": The model using both B-mode and CEUS ($BM+CEUS_R$) significantly outperformed the model using only B-mode. The temporal data from the contrast phases—capturing how blood moves through the tumor—is essential for identifying the MTM subtype's unique vascular architecture.
  • Radiomics Trumps Clinical Data: Interestingly, adding clinical indicators (like AFP levels or cirrhosis status) to the radiomics model ($Hybrid_{R+C}$) did not significantly improve performance. This suggests that the radiomic features extracted from the images already capture the underlying biological and pathological signals.
  • Beyond Diagnosis to Prognosis: The model didn't just name the subtype; it predicted the future. Patients the model flagged as "MTM-positive" had significantly lower recurrence-free survival (RFS) rates within one year, matching the actual clinical outcomes.

Key Results

  • Diagnostic Power: The $BM+CEUS_R$ model achieved an AUC of 0.880 in the validation set and 0.878 in the external test set.
  • Comparison: This was vastly superior to the conventional clinic-radiological model (AUC 0.59–0.65), which relies on human interpretation of standard features.
  • High Specificity: In the validation set, the model showed a specificity of 90.9%, meaning it was excellent at correctly identifying patients who did not have the aggressive subtype, helping avoid unnecessary alarm.
  • Prognostic Significance: The model successfully stratified patients into high-risk and low-risk groups for early recurrence ($p < 0.001$).

Limitations and Future Work

While promising, the authors acknowledge several hurdles:

  • Sample Bias: The MTM subtype is relatively rare (about 30% of their cohort), which can lead to bias in machine learning.
  • Geographic Specificity: The dataset consisted primarily of East Asian patients with a high prevalence of Hepatitis B-related HCC. The model needs further validation in Western populations where Hepatitis C or fatty liver disease (NAFLD/MASLD) are more common causes of cancer.
  • Retrospective Nature: The study looked at past data. A prospective trial—where the model is used to make real-time predictions before surgery—is the next logical step to prove its clinical utility.

Reference: Lu, D., Qin, C., Wang, LF. et al. B-mode ultrasound and contrast-enhanced ultrasound-based radiomics interpretable analysis for the prediction of macrotrabecular-massive subtype of hepatocellular carcinoma. The Ultrasound Journal (2025) 17:53. https://doi.org/10.1186/s13089-025-00452-2

Summary updated

Abstract

The AdaBoost-based BM + CEUS<sub>R</sub> model shows promise as a tool for preoperatively identifying MTM-HCC and may also be beneficial in predicting prognosis.
BibTeX
@article{lu2025b,
  author = {Dan Lu and Cheng Qin and Lifan Wang and Lingling Li and Yu Li and Liping Sun and Hui Shi and Boyang Zhou and Xin Guan and Miao Yao and Hong Han and Jianhua Zhou and Huixiong Xu and Chong-Ke Zhao},
  title = {{B-mode ultrasound and contrast-enhanced ultrasound-based radiomics interpretable analysis for the prediction of macrotrabecular-massive subtype of hepatocellular carcinoma}},
  journal = {The Ultrasound Journal},
  year = {2025},
  doi = {10.1186/s13089-025-00452-2},
}
Figure