Machine Learning Model Integrating Clinical Features, CT Radiomics, and Transfer Learning for Predicting Aggressive Recurrence of Hepatocellular Carcinoma

Technical Summary: Machine Learning Model Integrating Clinical Features, CT Radiomics, and Transfer Learning for Predicting Aggressive Recurrence of Hepatocellular Carcinoma

Problem Statement
Hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality, with post-curative hepatectomy recurrence rates reaching 60–70%. A critical clinical challenge is the prediction of "aggressive recurrence," defined as the first recurrence exceeding the Milan criteria (e.g., large tumor burden, vascular invasion, or extrahepatic metastasis). Current risk assessment relies heavily on postoperative pathological features (e.g., microvascular invasion) or serum markers like AFP, which lack sensitivity for preoperative decision-making. While radiomics offers a noninvasive alternative, traditional handcrafted features often struggle to capture complex spatial patterns of tumor aggressiveness and are prone to overfitting in small datasets. Furthermore, deep learning models typically require large annotated datasets, which are scarce in medical imaging. This study addresses the need for a preoperative, noninvasive prediction model that integrates clinical data, conventional radiomics, and deep learning features to identify high-risk patients for individualized perioperative strategies.

Methodology
The study employed a retrospective design involving 455 patients with HCC who underwent curative hepatectomy at the First Affiliated Hospital of Guangxi Medical University (2013–2020). The cohort was split into a training group (n=318) and a validation group (n=137) at a 7:3 ratio.

1. Data Acquisition and Preprocessing:

   • Clinical Data: Preoperative variables included demographics, laboratory tests (e.g., albumin, AFP, FIB-4), and imaging measurements (e.g., portal vein diameter).
   • Imaging: Preoperative portal venous phase contrast-enhanced CT images were used. Regions of Interest (ROIs) were manually segmented using ITK-SNAP by two senior physicians.
   • Radiomics: Using PyRadiomics, 1,561 handcrafted features were extracted, covering first-order statistics, texture matrices (GLCM, GLRLM, GLSZM, NGTDM, GLDM), and shape features. Features underwent Z-score normalization.
   • Deep Learning (Transfer Learning): A ResNet-50 model, pre-trained on ImageNet, was fine-tuned on the training set. Deep features were extracted from the global average pooling layer (2,048 dimensions) and reduced to 150 principal components via Principal Component Analysis (PCA).

2. Feature Selection and Model Construction:

   • Selection: Features were screened using univariate t-tests, correlation analysis (threshold 0.9), and Least Absolute Shrinkage and Selection Operator (LASSO) regression to eliminate redundancy and select the most predictive subset.
   • Multimodal Fusion: An "DLR" (Deep Learning and Radiomics) feature set was created via early fusion.
   • Algorithms: Eight machine learning algorithms (Logistic Regression, SVM, KNN, Random Forest, Extra Trees, XGBoost, LightGBM, MLP) were tested. XGBoost was selected as the optimal classifier based on performance.
   • Models Built: Four distinct models were constructed: Clinical-only, Radiomics-only, DLR-only, and a Multimodal Fusion model (Clinical + Radiomics + DLR).
   • Evaluation: Performance was assessed using the Area Under the Curve (AUC), calibration curves, and Decision Curve Analysis (DCA) on both training and independent validation groups.

Key Results

• Independent Predictors: Multivariate logistic regression identified four independent clinical factors for aggressive recurrence: Albumin (OR=0.927), FIB-4 score (OR=0.524), Portal Vein Diameter (OR=1.173), and involvement of liver segment S2 (OR=0.071).
• Training Performance: The Multimodal Fusion model achieved exceptional discriminative ability in the training group with an AUC of 0.996 (95% CI: 0.992–1.000). The DLR model also performed near-perfectly (AUC 0.994).
• Validation Performance: In the independent validation group, performance dropped significantly across all models, indicating severe overfitting. The Multimodal Fusion model achieved an AUC of only 0.567 (95% CI: 0.468–0.667). The DLR model dropped to 0.564, and the Clinical model to 0.436.
• Clinical Utility: Decision Curve Analysis (DCA) revealed that in the validation group, the net benefit of the Multimodal Fusion model did not exceed the "treat all" or "treat none" strategies across any threshold probability.

Significance and Claims
The paper claims to be the first to systematically integrate transfer learning-based deep features, conventional radiomics, and clinical variables for predicting aggressive HCC recurrence.

• Theoretical Feasibility: The study demonstrates that multi-source heterogeneous information fusion is theoretically feasible, as the combined model showed superior training performance compared to single-modality models, suggesting that deep learning features capture complementary information regarding tumor heterogeneity.
• Clinical Caution: Crucially, the authors explicitly state that the model is currently unsuitable for clinical decision-making. The significant gap between training and validation AUCs, coupled with the lack of net clinical benefit in DCA, confirms that the model suffers from overfitting due to the "high-dimensional features – limited samples – complex models" dilemma.
• Future Direction: The authors emphasize that the primary value of this work lies in providing a methodological reference for multimodal fusion strategies. They conclude that large-scale, multicenter, prospective studies with larger sample sizes are required to validate and optimize the model before it can be applied to individualized perioperative treatment planning.

In summary, while the study successfully constructs a high-performing model in a retrospective setting, it modestly concludes that its generalizability is limited, and it serves primarily as a proof-of-concept for integrating deep learning with radiomics in small-sample medical datasets rather than a ready-to-use clinical tool.