Multi-task deep learning network to predict future macrovascular invasion in hepatocellular carcinoma

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Predictive models for macrovascular invasion in hepatocellular carcinoma (HCC) are crucial for timely clinical interventions.
• Hepatocellular carcinoma (HCC) is a significant global health concern, with macrovascular invasion being a key factor influencing prognosis.

Purpose of the Study:

• To develop and validate a multi-task deep learning network-based model for predicting future macrovascular invasion in HCC.
• To compare the performance of different deep learning models and identify the optimal approach for prediction.

Main Methods:

• Retrospective collection of 366 HCC cases from five Chinese hospitals.
• Construction of multi-task deep learning models incorporating clinical/radiological factors and radiomic features.
• External validation of the best-performing model using a separate dataset.

Main Results:

• The optimal model, combining clinical/radiological factors and radiomic features, demonstrated high discrimination (AUC 0.877 training, 0.836 validation).
• The model's predictions were statistically significant for time to macrovascular invasion and overall survival in both training and validation datasets.
• Segmentation subnets and end-to-end deep learning algorithms enhanced model performance.

Conclusions:

• A multi-task deep learning network effectively predicts future macrovascular invasion in HCC.
• The model's robust performance suggests potential for improved risk stratification and personalized treatment strategies in high-risk HCC populations.