Accuracy of Medical Image-Based Deep Learning for Detecting Microvascular Invasion in Hepatocellular Carcinoma: Systematic Review and Meta-Analysis.

Background: Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide. Microvascular invasion (MVI) is a critical pathological indicator of postoperative recurrence and poor prognosis in patients with HCC. Some researchers have explored the diagnostic accuracy of deep learning (DL) based on various imaging modalities for MVI. Objective: This meta-analysis aimed to systematically evaluate the preoperative diagnostic performance of DL models using medical images to predict MVI in HCC, and to investigate the impact of different imaging modalities and validation strategies on model performance and generalizability. Methods: PubMed, Cochrane Library, Embase, and Web of Science were searched up to October 16, 2025. Studies investigating the detection of MVI in HCC using imaging-based DL techniques were eligible. Studies focusing solely on image segmentation were excluded. The Quality Assessment of Diagnostic Accuracy Studies-2 tool was used to assess risk of bias. A bivariate mixed-effects meta-analysis was performed to calculate the pooled sensitivity, specificity, and area under the summary receiver operating characteristic curve (SROC). Subgroup analyses were conducted by imaging modality and validation set generation method. Results: This meta-analysis included 52 studies with 19,531 patients with HCC. The pooled analysis revealed that imaging-based DL models had an overall sensitivity of 0.80 (95% CI 0.78-0.83), a specificity of 0.82 (95% CI 0.80-0.85), and an SROC of 0.88 for MVI prediction. Subgroup analysis showed that models based on preoperative contrast-enhanced computed tomography performed excellently, with a sensitivity of 0.84 (95% CI 0.79-0.88), a specificity of 0.83 (95% CI 0.77-0.88), and an SROC of 0.90. These results suggest that contrast-enhanced computed tomography is the most promising noninvasive method for current clinical applications. Meanwhile, DL models using pathological sections achieved the highest diagnostic performance: a sensitivity of 0.91 (95% CI 0.87-0.94), a specificity of 0.90 (95% CI 0.68-0.97), and an SROC of 0.92. This establishes the ultimate benchmark for performance optimization for all noninvasive models. A key finding was that model performance was less consistent in independent external validation (SROC: 0.85) than in internal validation (SROC: 0.90). This discrepancy indicates that overreliance on internal validation may overestimate model efficacy and underscores the decisive role of rigorous external validation in assessing real-world generalizability. Conclusions: This study is the first to systematically assess the use of imaging-based DL for diagnosing MVI in HCC. The results demonstrate a significant potential for these models in predicting MVI. However, their clinical applicability requires rigorous evaluation, given the scarcity of independent external validation cohorts, notable heterogeneity among them, and the observed decline in model performance. Therefore, prospective, multicenter studies following standardized reporting guidelines are a critical future direction. These studies should also focus on developing integrated algorithms that translate histopathological insights into preoperative imaging data to establish robust clinical tools.