Automatic classification of HEp-2 specimens by explainable deep learning and Jensen-Shannon reliability index

Area of Science:

• Immunology
• Computational Biology
• Medical Diagnostics

Background:

• The Anti-Nuclear Antibodies (ANA) test using Human Epithelial type 2 (HEp-2) cells via Indirect Immuno-Fluorescence (IIF) is the gold standard for diagnosing Connective Tissue Diseases.
• Computer-assisted analysis of HEp-2 images is crucial for advancing ANA pattern classification, leveraging machine learning.

Purpose of the Study:

• To develop an innovative platform for automated HEp-2 image analysis and ANA pattern classification.
• To enhance the accuracy, robustness, and explainability of computer-assisted ANA testing.

Main Methods:

• Utilized transfer learning with pre-trained deep learning models for HEp-2 image analysis.
• Implemented unsupervised deep description, a novel feature selection for unbalanced datasets, and independent validation on multi-hospital data.
• Introduced modified gradient-weighted class activation mapping for regional explainability and a Jensen-Shannon divergence-based sample quality index.

Main Results:

• Achieved exceptionally high performance in intensity and ANA pattern recognition, surpassing state-of-the-art methods.
• Demonstrated robustness and versatility by eliminating the need for cell segmentation, relying instead on statistical sample analysis.
• Validated the method's cross-hardware compatibility and reliability across different hospital datasets.

Conclusions:

• The developed platform offers a robust, versatile, and highly accurate solution for automated HEp-2 image analysis in ANA testing.
• The AI-driven approach significantly improves the classification of ANA patterns, aiding in the diagnosis of connective tissue diseases.
• Future work will extend the approach to recognize complex patterns, including mitotic spindle recognition and mixed patterns.