HP-DIL: Deep heterogeneity profiling with graph-informed disentangled interaction learning for MRI-based liver...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Liver fibrosis staging (LFS) is crucial for managing liver disease, but current methods rely on invasive biopsies.
• Multiparametric MRI offers non-invasive quantitative assessment of liver fibrosis.
• Deep learning for automated LFS faces challenges due to patient and regional heterogeneities and limited interaction modeling.

Purpose of the Study:

• To introduce a novel deep learning framework, heterogeneity profiling with graph-informed disentangled interaction learning (HP-DIL), for accurate and interpretable non-invasive liver fibrosis staging.
• To address limitations of existing deep learning models in handling etiological and tissue-level heterogeneities and inter-regional interactions.

Main Methods:

• HP-DIL employs unsupervised subregion discovery by fusing multiparametric MRI signals, spatial-texture, and anatomical priors to create subject-level graphs.
• A global-local graph transformer captures higher-order interactions within spatial-semantic interaction graphs.
• A disentangled interaction mechanism (DIM) based on causal inference decouples features and mitigates spurious correlations.

Main Results:

• HP-DIL demonstrated competitive accuracy and cross-center generalizability on two external test cohorts from multi-vendor centers.
• Qualitative analysis showed agreement between DIM-highlighted regions and radiological assessments, clarifying imaging relevance.
• The framework successfully mitigates spurious correlations while preserving disease-relevant signals.

Conclusions:

• HP-DIL provides an accurate and interpretable approach for non-invasive liver fibrosis staging using multiparametric MRI.
• The framework's ability to handle heterogeneity and improve generalizability supports its potential for clinical deployment.
• This study advances automated LFS by integrating graph-based learning and disentangled representation.