JMIR Publications

ABSTRACT

Background:

Microvascular invasion (MVI) represents a significant risk factor for recurrence and unfavorable prognosis following surgical resection in individuals with hepatocellular carcinoma (HCC). Accurate early identification of MVI status is of critical importance to clinicians in selecting appropriate treatment strategies and improving overall patient survival. In recent years, the incorporation of computer-aided diagnosis systems based on deep learning (DL) technology has provided new possibilities for improving MVI prediction accuracy. However, a systematic body of evidence concerning the diagnostic performance of different imaging modalities for HCC MVI remains lacking.

Objective:

The present meta-analysis was undertaken to systematically assess the diagnostic precision of image-based DL for MVI in HCC.

Methods:

The Cochrane Library, PubMed, Web of Science, and Embase were retrieved up to November 5, 2024 to assess the diagnostic efficacy of DL based on medical images in HCC MVI. The Quality Assessment of Diagnostic Accuracy Studies 2 tool was applied to assess the risk of bias in eligible studies. Meta-analysis was conducted using the validation dataset.

Results:

This meta-analysis incorporated 42 articles in total, encompassing 16,053 HCC subjects. The included studies were based on contrast-enhanced magnetic resonance imaging (CEMRI) (12 articles), contrast-enhanced computed tomography (CECT) (14 articles), contrast-enhanced ultrasound (CEUS) (three articles), pathological images (two articles), MRI (six articles), and multimodal imaging (CT+PET: one article, CECT+CEMRI: four articles). A summary of the DL model's diagnostic performance in predicting HCC MVI using medical images yielded a sensitivity of 0.81 (95% CI: 0.78, 0.84), a specificity of 0.82 (95% CI: 0.79, 0.85), and an SROC of 0.88 (95% CI: 0.56, 0.98). For CECT, the results demonstrated a sensitivity of 0.85 (95% CI: 0.79, 0.89), a specificity of 0.83 (95% CI: 0.76, 0.89), and an SROC of 0.91 (95% CI: 1.00, 0.00). For CEMRI, the results indicated a sensitivity of 0.81 (95% CI: 0.75, 0.86), a specificity of 0.79 (95% CI: 0.73, 0.85), and an SROC of 0.87 (95% CI: 1.00, 0.00). Models based on pathological sections exhibited the highest diagnostic efficacy, with a sensitivity, specificity, and SROC of 0.91 (95% CI: 0.87, 0.94), 0.90 (95% CI: 0.68, 0.97), and 0.92 (95% CI: 1.00, 0.00), respectively. According to the subgroup analysis based on the validation set generation method, the sensitivity, specificity, and SROC were 0.82 (95% CI: 0.78, 0.86), 0.83 (95% CI: 0.79, 0.87), and 0.90 (95% CI: 1.00, 0.00) in internal validation, and 0.80 (95% CI: 0.74, 0.84), 0.78 (95% CI: 0.72, 0.83), and 0.86 (95% CI: 1.00, 0.00) in external validation, respectively.

Conclusions:

Medical image-based DL models demonstrate good accuracy in the preoperative noninvasive diagnosis of MVI in HCC. In addition, the current study discovered that DL models based on invasive imaging of pathological sections exhibited higher diagnostic efficacy. Clinical Trial: PROSPERO, CRD42024613733