Introduction
Hepatocellular Carcinoma (HCC) remains a primary cause of cancer-related mortality worldwide. In clinical practice, diagnosis and staging rely on a synthesis of diverse data sources, including blood work, patient history, and multiple imaging modalities such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). Despite this, machine learning research in oncology often operates in silos, focusing on either image-based radiomics or tabular clinical data independently. This study addresses this gap by establishing a multi-modal machine-learning baseline for HCC diagnosis, specifically targeting TNM staging—a critical metric for describing cancer spread.
Methodology
The researchers developed a comprehensive pipeline to integrate heterogeneous data types for classification using the XGBoost algorithm:
- Radiomics Extraction: Using the
PyRadiomicslibrary, features were extracted from segmented lesions in contrast-enhanced CT and MRI scans. These features capture tumor texture, shape, and intensity distributions often imperceptible to human clinicians. - Tabular Data Processing: Clinical data from REDCap case report forms and laboratory results were digitized. Missing values were addressed via mean imputation to maintain dataset integrity.
- Data Augmentation: To maximize the utility of the 100-patient cohort, the authors generated 376 unique samples by pairing various combinations of a patient’s available MRI and CT scans with their corresponding laboratory data.
- Feature Selection: Mutual Information and Recursive Feature Elimination (RFE) were employed to identify the most predictive variables and mitigate the curse of dimensionality.
Dataset: The QHCC Cohort
The study introduces the Quality Hepatocellular Carcinoma (QHCC) dataset, comprising 100 patients. The dataset includes contrast-enhanced T1w MRI, contrast-enhanced CT, clinical laboratory reports, and REDCap forms. Ground truth for the classification task was derived from 95 patients with confirmed TNM staging labels.
Results and Discussion
The experimental results underscore the necessity of multi-modal integration. The XGBoost model achieved its peak performance when combining all available data sources (CT + MRI + REDCap + Lab), reaching an Accuracy (ACC) of 0.89 ± 0.05 and an Area Under the Curve (AUC) of 0.93 ± 0.03.
In contrast, single-modality models performed significantly worse. For instance, using MRI data alone resulted in an ACC of only 0.51. Furthermore, the inclusion of laboratory data alongside clinical history (REDCap) improved the AUC from 0.88 to 0.90, demonstrating that active blood work provides complementary signals to static patient history.
Feature Importance and Interpretability
The model's predictive logic aligns closely with established clinical intuition. Analysis of feature importance scores reveals that the most influential tabular features are "Tumor Value" (representing size and count) and "Alkaline Phosphatase" (a key liver enzyme). In the radiomics domain, the model prioritized the mean intensity and kurtosis of the MRI signal.
Limitations and Future Work
While providing a strong baseline, the study notes several limitations. The original cohort of 100 patients is relatively small, necessitating data augmentation. Additionally, class imbalance in T-stage labels required the merging of categories (e.g., "T3 or T4") to ensure model feasibility. Finally, the reliance on manual lesion segmentation by medical professionals suggests a future opportunity for end-to-end automation using deep learning architectures like U-Net.
Summary updated