For patients with hepatocellular carcinoma (HCC), the presence of microvascular invasion (MVI) is one of the most important prognostic factors. Tumor cells that have infiltrated nearby blood vessels dramatically increase the risk of recurrence after surgical resection or ablation. But MVI can currently only be confirmed after surgery, by examining the resected tissue under a microscope, too late to inform the initial treatment decision.
A deep learning model developed across 23 hospitals in China may change that. Called MAPUSE (MVI AI Prediction via Contrast-enhanced Ultrasound with Explainability), the model predicts MVI from pre-operative contrast-enhanced ultrasound (CEUS) videos, achieving AUC values between 0.835 and 0.978 across multiple validation cohorts.
The study, published July 10 in Nature Communications (DOI: 10.1038/s41467-026-74985-y), was led by researchers at the Chinese PLA General Hospital, the Institute of Automation of the Chinese Academy of Sciences, and 21 other centers.
A Transformer for Ultrasound
The MAPUSE model uses a TimeSformer architecture, a fully Transformer-based video model with no convolutional backbone, to analyze 48 frames from each CEUS video, sampled across the arterial (0-45 seconds), venous (45-120 seconds), and delay (180-190 seconds) phases. The region of interest includes a 1.2-times enlarged bounding box around the tumor plus peritumoral tissue, which outperformed tumor-only analysis.
The team assembled 5,148 CEUS videos from 1,716 HCC patients across 23 hospitals, one of the largest ultrasound-AI datasets assembled for this purpose. The videos comprised more than 3.3 million individual frames. Training used 495 patients, with internal validation on 213, hold-out testing on 190 patients using two different ultrasound contrast agents (Sonovue and Sonazoid), and prospective validation in two independent cohorts, one in southern China (Guangzhou) and one in northern China (Beijing).
Performance Across Cohorts
The model’s AUC ranged from 0.835 on the hold-out Sonovue test set to 0.986 on the training set. In prospective validation, the southern cohort achieved an AUC of 0.886 and the northern cohort 0.847.
Performance was strongly dependent on tumor size. For tumors larger than 5 cm, the AUC reached 0.978. For tumors between 3 and 5 cm, it dropped to 0.814, and for tumors smaller than 3 cm, precisely the group where pre-operative MVI assessment would be most clinically valuable, the AUC fell to 0.756.
Biological Explainability
One of the study’s notable contributions is its attempt to explain what the model is actually detecting. Only 15% of the model’s high-attention areas on ultrasound heatmaps coincided with the actual physical location of microvascular invasion, meaning the model is not directly seeing tumor emboli in blood vessels. Instead, through a triple-verification analysis involving bulk RNA sequencing (203 patients), single-cell RNA sequencing (12 patients, 86,412 immune cells), and immunohistochemistry (64 patients, 160 attention areas), the team showed that high MVI risk scores correlate with reduced CD8+ T cell infiltration in the tumor microenvironment.
“MVI high-risk patients have an immune desert phenotype,” the authors write. The model appears to detect ultrasound features that correlate with immune exclusion, a tumor microenvironment hostile to cytotoxic T cells, which in turn is associated with higher likelihood of vascular invasion.
Clinical Implications
In a separate cohort of 568 ablation patients, those classified as MVI high-risk by MAPUSE who received adjuvant immunotherapy had significantly better 5-year disease-free survival than those who did not (30.8% vs. 14.6%; hazard ratio 0.61). This suggests the model could help identify patients most likely to benefit from immunotherapy after ablation, a group currently selected using criteria that do not include MVI status.
The authors emphasize that the model is not intended to replace pathology but to provide a pre-operative risk assessment that currently does not exist. MVI status remains a post-surgical finding; MAPUSE offers a window into that information before the first incision.
Source: Pang, C., Ru, J., Liu, Y. et al. “Prediction of microvascular invasion in hepatocellular carcinoma using contrast-enhanced ultrasound and deep learning.” Nature Communications (2026). DOI: 10.1038/s41467-026-74985-y