Modality-Aware Distillation Network for Microvascular Invasion Prediction of Hepatocellar Carcinoma From MRI Images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Microvascular invasion (MVI) in hepatocellular carcinoma (HCC) is a key predictor of recurrence after liver surgery or transplantation.
• Combining clinicoradiologic data with medical images improves HCC prediction, but clinical data is often difficult to obtain.
• There is a need for methods that can leverage existing clinical data knowledge to enhance HCC classification using only imaging data.

Purpose of the Study:

• To develop a modality-aware distillation network (MD-Net) for transferring knowledge from clinical data to an image-only student network.
• To improve HCC classification accuracy by enabling the student network to utilize transferred clinical information.
• To enhance the student network's performance through a novel self-supervised task for predicting clinical data from images.

Main Methods:

• A teacher network integrates 3D MRI images and non-image clinicoradiologic data using fusion and attention modules.
• A student network processes 3D MRI data using modality-specific modules and an attention mechanism.
• Classification-level and feature-level distillation techniques transfer knowledge from the teacher to the student network.
• A self-supervised task predicts clinical characteristics from imaging data to further boost classification.

Main Results:

• The proposed MD-Net effectively transfers clinical information to the image-only student network.
• The method demonstrated improved HCC classification performance.
• Achieved AUC scores of 71.86% and 75.51% on independent datasets, outperforming state-of-the-art methods.

Conclusions:

• Knowledge distillation from clinical and image data to an image-only network is effective for HCC classification.
• The MD-Net architecture and self-supervised task enhance diagnostic capabilities for HCC.
• This approach offers a promising solution for improving HCC prediction when comprehensive clinical data is unavailable.