IVIM using convolutional neural networks predicts microvascular invasion in HCC

Area of Science:

• Radiology and Imaging Science
• Artificial Intelligence in Medicine
• Oncology and Hepatobiliary Diseases

Background:

• Microvascular invasion (MVI) is a critical prognostic factor in hepatocellular carcinoma (HCC).
• Accurate preoperative prediction of MVI is essential for guiding treatment decisions in HCC management.
• Intravoxel incoherent motion (IVIM) diffusion-weighted magnetic resonance imaging (DWI) shows potential for assessing tissue microstructure.

Purpose of the Study:

• To evaluate the diagnostic performance of IVIM-DWI using convolutional neural networks (CNNs) for predicting MVI in HCC.
• To compare the performance of deep learning models based on IVIM data with traditional IVIM parameter maps and clinical features.
• To develop and assess a fusion model combining deep IVIM features, clinical characteristics, and IVIM parameters for enhanced MVI prediction.

Main Methods:

• Retrospective analysis of 114 HCC patients with preoperative IVIM-DWI MRI (9 b-values).
• Development of a CNN model to extract deep features directly from IVIM b-value volumes.
• Construction of models based on IVIM parameter maps, a fusion of deep IVIM features with clinical data (AFP, tumor size) and ADC, and ROC analysis for performance evaluation.

Main Results:

• The CNN model extracting deep features directly from IVIM-DWI demonstrated superior performance (AUC 0.810) compared to models based on IVIM parameter maps (AUC 0.590) for MVI prediction.
• A fusion model incorporating deep IVIM features, clinical characteristics (AFP, tumor size), and ADC achieved a slightly improved AUC of 0.829.
• Deep learning assessment of IVIM data overcomes the limitations of unstable and low performance associated with traditional IVIM parameters.

Conclusions:

• Deep learning utilizing CNNs with IVIM-DWI is a promising tool for the preoperative prediction of MVI in HCC.
• The deep learning approach based on IVIM data outperforms models relying solely on IVIM parameter maps or clinical features.
• The developed fusion model offers enhanced diagnostic performance for MVI prediction in HCC patients.