Quantitative benchmarking of anomaly detection methods in digital pathology images

Area of Science:

• Computer Science
• Medical Imaging
• Pathology

Background:

• Anomaly detection is crucial for industrial defect inspection.
• Digital pathology offers potential applications like rare disease identification and biomarker discovery.
• Pathology images present unique challenges (large size, multi-scale, stain variability) that hinder current anomaly detection algorithms.

Purpose of the Study:

• To quantitatively benchmark 23 classical anomaly detection methods for digital pathology.
• To evaluate the influence of image scale, anomaly patterns, and training strategies on performance.
• To provide a comprehensive comparison of method strengths and limitations for future research.

Main Methods:

• Extensive experiments were conducted on five digital pathology datasets (real and synthetic).
• 23 classical anomaly detection algorithms were systematically evaluated.
• Performance was analyzed based on image scale, anomaly pattern types, and training epoch selection.

Main Results:

• A detailed comparison of the strengths and limitations of each benchmarked method was established.
• Significant variations in detection performance were observed across different methods and experimental conditions.
• The study provides a foundational benchmark for anomaly detection in digital pathology.

Conclusions:

• Current anomaly detection algorithms face challenges with unique digital pathology image characteristics.
• The benchmark established in this study will inform the selection and development of future algorithms.
• Publicly available code and data will facilitate reproducibility and further research in pathology anomaly detection.