Improved diagnostic decision making for microvascular invasion in HCC using a novel nomogram incorporating delta radiomics and body composition factors: A multicenter study

Abstract

Objective

To develop and validate machine learning(ML) models based on delta-radiomics features and body composition factors for early prediction of microvascular invasion(MVI) in patients with hepatocellular carcinoma(HCC) using a multicenter cohort,and to identify differentially expressed genes(DEGs).

Methods

This retrospective study included pathologically-confirmed HCC patients diagnosed at three centers.Radiomic features were extracted from MRI images,and delta-radiomics features were calculated.Clinical-radiological features, body composition factors and delta-radiomics score were selected through various feature selection methods and a nomogram was built based on the independent risk factors.The performance of the nomogram was assessed with the area under the receiver operating characteristic curve (AUC).Recurrence-free survival(RFS) analysis was assessed by the Kaplan-Meier analysis and compared using the log-rank test.Additionally, gene expression analysis was conducted to explore molecular mechanisms underlying MVI.

Results

The nomogram demonstrated numerically superior predictive performance in both external test sets, achieving AUCs of 0.853 (test set1) and 0.852 (test set2). The Delong test revealed the nomogram demonstrated robust predictive performance across both external test set, compared to the clinical model (test set1: 0.853 vs 0.790; test set2: 0.852 vs 0.774; both p < 0.05). No statistically significant difference was observed between the nomogram and delta-radiomics model(p > 0.05).The nomogram's implementation enhanced radiologists' diagnostic accuracy for MVI by up to 13.4 percentage points.The nomogram can categorize recurrence-free survival.DEGs associated with MVI are related to cell proliferation and glucose metabolism.

Conclusion

The ML models established via body composition factors and delta-radiomics scores had the best performance to predict MVI status,and help improve the diagnostic capability of radiologists.