Diagnostic Accuracy of Artificial Intelligence Based on Imaging Data for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma: A Systematic Review and Meta-Analysis.

BACKGROUND

The presence of microvascular invasion (MVI) is considered an independent prognostic factor associated with early recurrence and poor survival in hepatocellular carcinoma (HCC) patients after resection. Artificial intelligence (AI), mainly consisting of non-deep learning algorithms (NDLAs) and deep learning algorithms (DLAs), has been widely used for MVI prediction in medical imaging.

AIM

To assess the diagnostic accuracy of AI algorithms for non-invasive, preoperative prediction of MVI based on imaging data.

METHODS

Original studies reporting AI algorithms for non-invasive, preoperative prediction of MVI based on quantitative imaging data were identified in the databases PubMed, Embase, and Web of Science. The quality of the included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) scale. The pooled sensitivity, specificity, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were calculated using a random-effects model with 95% CIs. A summary receiver operating characteristic curve and the area under the curve (AUC) were generated to assess the diagnostic accuracy of the deep learning and non-deep learning models. In the non-deep learning group, we further performed meta-regression and subgroup analyses to identify the source of heterogeneity.

RESULTS

Data from 16 included studies with 4,759 cases were available for meta-analysis. Four studies on deep learning models, 12 studies on non-deep learning models, and two studies compared the efficiency of the two types. For predictive performance of deep learning models, the pooled sensitivity, specificity, PLR, NLR, and AUC values were 0.84 [0.75-0.90], 0.84 [0.77-0.89], 5.14 [3.53-7.48], 0.2 [0.12-0.31], and 0.90 [0.87-0.93]; and for non-deep learning models, they were 0.77 [0.71-0.82], 0.77 [0.73-0.80], 3.30 [2.83-3.84], 0.30 [0.24-0.38], and 0.82 [0.79-0.85], respectively. Subgroup analyses showed a significant difference between the single tumor subgroup and the multiple tumor subgroup in the pooled sensitivity, NLR, and AUC.

CONCLUSION

This meta-analysis demonstrates the high diagnostic accuracy of non-deep learning and deep learning methods for MVI status prediction and their promising potential for clinical decision-making. Deep learning models perform better than non-deep learning models in terms of the accuracy of MVI prediction, methodology, and cost-effectiveness.

SYSTEMATIC REVIEW REGISTRATION

https://www.crd.york.ac.uk/PROSPERO/display_record.php? RecordID=260891, ID:CRD42021260891.