Learning ordinal representation across MRI sequences for liver fibrosis staging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate liver fibrosis staging (LFS) is crucial for managing liver diseases.
• Noncontrast MRI is ideal for screening and monitoring but lacks detail for precise staging.
• Further exploration of noncontrast MRI for LFS is needed due to fibrosis progression and subtle stage differences.

Purpose of the Study:

• Develop and evaluate a fine-grained deep learning pipeline for precise LFS using noncontrast MRI.
• Improve diagnostic accuracy for liver diseases through advanced AI techniques.

Main Methods:

• A fine-grained diagnostic model was developed using noncontrast MRI from 450 cases.
• Multi-scale learning and attention mechanisms were employed to enhance information utilization.
• A hybrid contrastive triplet learning method with a weighted strategy addressed data imbalance and improved performance.

Main Results:

• The fine-grained model achieved an AUC of 0.877, surpassing existing LFS and general MRI models.
• AUC increased to 0.930 for identifying cirrhosis.
• UMAP clustering and gradient-based methods revealed predictive mechanisms, showing potential for fine-grained diagnosis.

Conclusions:

• The fine-grained deep learning model effectively uses multisequence noncontrast MRI for precise LFS.
• The model demonstrates improved diagnostic performance for liver fibrosis staging.