Differentiating advanced from non-advanced hepatic fibrosis: a hybrid deep learning-radiomics model leveraging...

This study aimed to develop and validate a hybrid deep learning-radiomics model that leveraged Cycle-consistent generative adversarial networks (CycleGAN)-synthesized contrast-enhanced computed tomography (CE-CT) images to differentiate advanced from non-advanced hepatic fibrosis.

This retrospective study included 410 patients with biopsy-confirmed hepatic fibrosis (2017-2024). A trained CycleGAN model was used to generate synthetic three-phase CE-CT images from the corresponding non-contrast computed tomography (NC-CT) data. Each group of images was randomly split 6:4 ratio into training and test sets. After region of interest (ROI) segmentation, handcrafted radiomic (HCR) features were extracted. Concurrently, eight end-to-end deep learning (DL) models were trained; DL features were extracted from the best-performing model. Feature selection was performed using Spearman's rank correlation and the least absolute shrinkage and selection operator (LASSO). Six machine learning classifiers were developed for each feature type (HCR, DL, and late-fused DL features) using the final selected feature set. The performance of models was assessed by the area under the receiver operating characteristic curve (AUC), accuracy, calibration curves, decision curve analysis (DCA) and the DeLong test.

Models utilizing synthetic CE-CT images outperformed those based on NC-CT. DL feature-based models surpassed HCR-based models. A late-fusion hybrid model integrating DL features further improved performance, achieving an AUC of 0.880 (95% CI: 0.819-0.942).

The model based on synthetic CE-CT images demonstrated excellent diagnostic performance. Moreover, the hybrid model combining both real NC-CT and synthetic CE-CT images further improved diagnostic performance. The hybrid model can serve as a non-invasive diagnostic method for differentiating advanced from non-advanced hepatic fibrosis.