Predicting microvascular invasion in hepatocellular carcinoma with a CT- and MRI-based multimodal deep learning model
Yan Lei1,2, Bao Feng3, Meiqi Wan1,2
1Department of Radiology, Jiangmen Central Hospital, 23 Beijie Haibang Street, Jiangmen, People's Republic of China.
Abdominal Radiology (New York)
|March 3, 2024
Summary
A multimodal deep learning (MDL) model using CT and MRI effectively predicts microvascular invasion (MVI) in hepatocellular carcinoma (HCC). This noninvasive approach shows superior performance compared to single-modality models and clinical features alone.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide.
- Microvascular invasion (MVI) is a critical prognostic factor in HCC, influencing treatment decisions and patient outcomes.
- Accurate preoperative prediction of MVI remains challenging with conventional methods.
Purpose of the Study:
- To evaluate the efficacy of a multimodal deep learning (MDL) model integrating computed tomography (CT) and magnetic resonance imaging (MRI) for predicting MVI in HCC.
- To compare the predictive performance of the MDL model against single-modality deep learning models and clinical factors.
Main Methods:
- Development of single-modality deep learning models using CT and MRI data from 235 HCC patients.
- Construction of an MDL model incorporating transfer learning (TL) with features from DenseNet121 and an extreme learning machine (ELM) classifier.
- Validation of the MDL model on a separate cohort of 110 HCC patients with simultaneous CT and MRI data.
Main Results:
- The MDL model achieved a superior area under the curve (AUC) of 0.844 compared to single-modality deep learning models (AUCs ranging from 0.722 to 0.731) and clinical features (AUC=0.648).
- The MDL model's performance surpassed that of individual CT and MRI models, demonstrating its enhanced predictive capability.
- Combining the MDL model with clinical features further improved its predictive accuracy (AUC=0.871), and a nomogram incorporating deep learning signatures showed significant clinical utility.
Conclusions:
- The developed MDL model serves as a valuable noninvasive tool for the preoperative prediction of MVI in HCC.
- Integration of multi-imaging data through deep learning significantly enhances the accuracy of MVI prediction in HCC.
- The MDL model holds promise for improving risk stratification and guiding treatment strategies for HCC patients.
