A Transformer-Based Deep Learning Model for predicting Early Recurrence in Hepatocellular Carcinoma After Hepatectomy Using Intravoxel Incoherent Motion Images

Background: This study aimed to develop and validate a transformer framework-based deep learning (DL) network using intravoxel incoherent motion (IVIM) diffusion-weighted imaging (DWI) to predict early recurrence in hepatocellular carcinoma (HCC). Materials and Methods: This retrospective study included 122 patients with HCC who underwent magnetic resonance imaging examination, including an IVIM-DWI sequence with nine b-values, before resection. These were divided into training (n=85) and test (n=37) sets. A vision transformer (ViT) framework-based DL was developed to predict early recurrence in HCC. Deep features were extracted from nine b-value DWI images and IVIM parametric maps and fused to construct the fused DL (ViT-fDL) prediction model. A clinical model was constructed using multivariate logistic regression analysis. A combined model was constructed using deep features from the ViT-fDL model and clinical independent features. The performances of the models were evaluated by discrimination, calibration, and clinical applicability. Results: Among 122 patients (108 males,14 females; mean age, 51.0 ± 11.9 years), 49 (40.1%) experienced early recurrence. The respective areas under the curve for the training and test sets were 0.755 (95% Confidence interval (CI), 0.650– 0.842) and 0.764 (95% CI, 0.596– 0.887) using the clinical model, 0.968 (95% CI, 0.905– 0.994) and 0.815 (95% CI, 0.653– 0.923) using the ViT-fDL model, and 0.991 (95% CI, 0.940– 1.000) and 0.821 (95% CI, 0.660– 0.927) using the combined model. Conclusion: The ViT-fDL model based on IVIM can be useful for preoperative prediction early recurrence in HCC. The combined model was a more effective and precise prediction tool than other models, promising to guide individualized postoperative monitoring. Keywords: hepatocellular carcinoma, early recurrence, deep learning, vision transformer