LI-RADS-based hepatocellular carcinoma risk mapping using contrast-enhanced MRI and self-configuring deep learning
Róbert Stollmayer1,2, Selda Güven3, Christian Marcel Heidt4
1Clinic for Diagnostic and Interventional Radiology (DIR), Heidelberg University Hospital, Heidelberg, Germany. robert.stollmayer@med.uni-heidelberg.de.
Summary
Deep learning models using nnU-Net can automate hepatocellular carcinoma (HCC) risk assessment with gadoxetate disodium-enhanced MRI. This approach shows high detection performance for LR-5 lesions, aiding in clinical practice.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Hepatocellular carcinoma (HCC) diagnosis relies on gadoxetate disodium-enhanced MRI (EOB-MRI).
- Liver Imaging Reporting and Data System (LI-RADS) standardization improves interpretation but is complex and time-consuming.
- Deep learning, specifically nnU-Net, offers potential to streamline EOB-MRI analysis for HCC.
Purpose of the Study:
- To develop and evaluate an automated segmentation model for HCC risk stratification using nnU-Net.
- To assess the model's performance in LI-RADS v2018 classification.
Main Methods:
- Retrospective analysis of 602 HCC-risk patients with EOB-MRI examinations.
- Training of U-Net models using the nnU-Net framework for automatic segmentation.
- Evaluation of lesion detection, LI-RADS classification, and segmentation metrics on internal and external test sets.
Main Results:
- The model achieved high detection sensitivities and positive predictive values for LI-RADS lesions, particularly LR-5.
- F1 scores for LI-RADS classification and Sørensen-Dice coefficients for segmentation demonstrated robust performance, especially for LR-5 lesions.
Conclusions:
- Automated tumor risk maps generated by deep learning segmentation tools show high detection performance for LR-5 HCC lesions.
- Further multi-center studies are recommended to enhance automatic LI-RADS classification before clinical implementation.
