Multi-scale deep learning models based on MRI for predicting pathological differentiation and evaluating its
1Department of Gastroenterology, The Third Affiliated Hospital of Soochow University, Changzhou, China.
Background:
Pathological differentiation is a critical prognostic indicator of biological behavior in hepatocellular carcinoma (HCC). The aim of the study was to develop multi-scale deep learning (DL) models based on magnetic resonance imaging (MRI) for predicting pathological differentiation and its association with recurrence-free survival (RFS) in HCC.
Methods:
A cohort of 292 patients with HCC was included and randomly assigned to a training set (TS) (n=204) and a validation set (VS) (n=88). DL models, including 2-dimensional (2D), 2.5-dimensional (2.5D), and 3-dimensional (3D), were trained by the ResNet50 network and developed using the eXtreme Gradient Boosting (XGBoost) classifier. The performance of these multi-scale DL models in predicting poorly-differentiated HCC (pdHCC) was evaluated using area under the curve (AUC). The SHapley Additive exPlanations (SHAP) method was applied to interpret the optimal DL models.
Results:
The 2.5D model based on MRI achieved the highest AUC value of 0.91 [95% confidence interval (CI): 0.87-0.95] and 0.86 (95% CI: 0.58-1.00) for the prediction of pdHCC in the TS and VS. This outperformed both the 2D (AUC =0.88 and 0.84) and 3D (AUC =0.83 and 0.64) models. Additionally, cases predicted as pdHCC by our developed MRI2.5D model demonstrated significantly lower RFS values compared to non-pdHCC cases (25 vs. 50 months, P=0.006). The SHAP approach highlighted the weighted importance of DL features, providing insightful interpretation within the MRI2.5D model for predicting pdHCC.
Conclusions:
The MRI2.5D model demonstrated superior capability for predicting pathological differentiation and its association with RFS in HCC, serving as a valuable tool for treatment decision-making in patients with HCC.