Hybrid Deep Learning Integrating Multimodal CEUS and Enhanced MRI to Optimize Early-Stage HCC Treatment Decisions (HDL-CEUS-MRI)

Core ObjectiveTo develop a hybrid deep learning model based on contrast-enhanced ultrasound (CEUS) clinical data for predicting the early recurrence (ER) probability within 2 years after liver resection (LH) or microwave ablation (MWA) in patients with early-stage liver cancer (CNLC I/IIa), providing a basis for individualized treatment decisions.Key MethodsStudy DesignType: Multicenter retrospective cohort study (2009.10-2026.12, 30 hospitals)Population:Inclusion: First diagnosed HCC (single lesion ≤5cm or 2-3 lesions with the largest ≤3cm), Child-Pugh A/B grade, CEUS/MRI examination within one month before surgery.Exclusion: Extrahepatic metastasis/other malignancies, previous treatment history, missing imaging/follow-up data.Sample size: 1441 cases (LH group 609 cases, MWA group 831 cases).Data CollectionClinical Data: Age, hepatitis type, cirrhosis, tumor characteristics (size/location/number), laboratory indicators (AFP, etc.).Imaging Data:CEUS: Dynamic acquisition of arterial phase (AP), portal venous phase (PVP), and delayed phase (LP) videos after SonoVue injection.Enhanced MRI: Acquisition of AP, PVP, transitional phase (TP), and hepatobiliary-specific phase (HBP) images after Gd-EOB-DTPA enhancement.Follow-Up: Imaging re-examination (CEUS/CT/MRI) every 3-6 months within 2 years after surgery, with endpoints being recurrence/death/last follow-up (≥24 months).Model ConstructionInput Data:Imaging ROI: Manual drawing of the tumor and a 5mm peritumoral area, extracting 64×64 pixel frames (CEUS videos downsampled to 32 frames).Fusion Methods: Clinical CEUS, clinical MRI, and clinical CEUS MRI.Deep Learning Architecture:Feature Extraction: 2D-CNN (convolutional layers → spatial features) → Bi-LSTM (fusion of spatiotemporal features) → Attention module.Data Augmentation: Translation and flipping to reduce overfitting.Training/Validation:5-fold cross-validation, dividing the training/validation set in a 4:1 ratio (training LH and MWA groups separately).Optimal Model Selection: The one with the highest AUC.Decision ApplicationHybrid model: To predict the ER probabilities for LH or MWA treatment for the same patient separately.Treatment Recommendations:If the predicted ER probability: MWA > LH → recommend LH;If the predicted ER probability: LH ≥ MWA → recommend MWA.