Triplet-branch network with contrastive prior-knowledge embedding for disease grading.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Computational Pathology

Background:

• Accurate disease grading is essential for determining appropriate clinical interventions, ranging from observation to immediate surgery.
• Class imbalance in medical datasets poses a significant challenge for developing robust grading models.
• Existing methods may not fully leverage prior knowledge relevant to disease progression and grading.

Purpose of the Study:

• To develop an accurate and robust disease grading system that addresses class imbalance.
• To integrate grade-related prior knowledge into a deep learning framework.
• To improve clinical decision-making by providing reliable disease grade assessments.

Main Methods:

• Proposed a novel Triplet-Branch Network with ContRastive priOr-knoWledge embeddiNg (TBN-CROWN).
• Implemented three branches for representation learning, classifier learning, and grade-related prior-knowledge learning.
• Introduced a contrastive embedding module to embed prior knowledge via contrastive learning, addressing class imbalance.

Main Results:

• TBN-CROWN effectively handles class-imbalanced training samples.
• The model achieves satisfactory grading accuracy across diverse datasets and diseases.
• Demonstrated superior performance in grading diseases like fatigue fracture, ulcerative colitis, and diabetic retinopathy.

Conclusions:

• TBN-CROWN offers a promising approach for accurate disease grading, particularly in scenarios with imbalanced data.
• The integration of prior knowledge through contrastive learning enhances model robustness and accuracy.
• This method has the potential to significantly aid physicians in making timely and appropriate treatment decisions.