Liver fibrosis automatic diagnosis utilizing dense-fusion attention contrastive learning network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Liver fibrosis is a major health concern with high mortality.
• Diffusion-Weighted Imaging (DWI) is a non-invasive tool for liver fibrosis detection.
• Deep learning aids in analyzing DWI for fibrosis staging but faces challenges with limited data and parameter synergy.

Purpose of the Study:

• To develop a Multi-view Contrastive Learning Network for automated classification of multi-parameter DWI images.
• To explore potential synergies and cross-comparisons among different DWI parameters for enhanced diagnostic capabilities.

Main Methods:

• A Dense-fusion Attention Contrastive Learning Network (DACLN) was designed for DWI image recognition.
• A multi-view contrastive learning framework was employed to train and extract features from multi-parameter DWI data.
• A Dense-fusion module was integrated for feature combination and final label prediction.

Main Results:

• The DACLN model achieved high performance on clinical data, with average scores for accuracy, precision, recall, specificity, and F-1 score ranging from 0.87 to 0.89.
• Interpretability analysis using Grad-CAM and annotation confirmed the model's effectiveness.
• Key DWI parameters including IVIM-f, CTRW-β, and MONO-ADC demonstrated significant recognition ability and complementarity.

Conclusions:

• The developed method achieves competitive accuracy in liver fibrosis diagnosis, even with limited multi-parameter DWI datasets.
• Identified sensitive DWI parameters offer promising avenues for future research in liver fibrosis diagnosis.
• The approach highlights the potential of multi-view contrastive learning in medical image analysis.