Toward efficient slide-level grading of liver biopsy via explainable deep learning framework.

Area of Science:

• Medical image analysis
• Computational pathology
• Hepatology

Background:

• Chronic liver diseases require precise diagnosis for effective management.
• Traditional histological analysis has limitations.
• Existing deep learning models for liver biopsy analysis need improvement.

Purpose of the Study:

• To develop a novel deep learning framework for enhanced grading accuracy and interpretability of liver biopsies.
• To improve the diagnosis and treatment of chronic liver diseases.

Main Methods:

• A patch-level classification model with multi-scale feature extraction and fusion was developed.
• A slide-level aggregation framework was introduced to integrate local histological information.
• The models were trained and validated on 1322 liver biopsy cases across various staining methods.

Main Results:

• The slide-level model achieved high F1 scores (0.9 for inflammatory activity and steatosis) and rapid analysis (<1 minute/slide).
• The patch-level model showed strong performance (F1 score 0.64 for ballooning, 0.99 for other indicators) and transferability to public datasets.
• The framework demonstrated robust interpretability.

Conclusions:

• The proposed analytical framework provides a reliable basis for diagnosing and treating chronic liver diseases.
• The AI models offer practical utility in clinical settings due to their accuracy, speed, and interpretability.