Diffusion models for out-of-distribution detection in digital pathology.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Detecting anomalies or out-of-distribution (OOD) data is crucial for deploying machine learning in medical imaging.
• Unsupervised learning offers a promising, annotation-free approach to OOD detection.
• Denoising diffusion probabilistic models (DDPMs) have advanced unsupervised OOD detection.

Purpose of the Study:

• To benchmark unsupervised OOD detection methods in digital pathology using AnoDDPMs.
• To evaluate the performance of AnoDDPMs on whole-slide image analysis in the Camelyon16 challenge.
• To assess the capability of AnoDDPMs in identifying anomalous data in medical images.

Main Methods:

• Applied AnoDDPMs, based on DDPMs, utilizing fast sampling techniques for large-scale whole-slide image analysis.
• Conducted patch-level OOD detection tasks on the Camelyon16 challenge dataset.
• Evaluated model performance using Receiver Operating Characteristic (ROC) analysis and Area Under the Curve (AUC) metrics.

Main Results:

• AnoDDPMs achieved high AUC scores of up to 94.13% and 86.93% on two patch-level OOD detection tasks.
• AnoDDPMs outperformed other unsupervised methods in detecting OOD data.
• Observed that AnoDDPMs modify anomalous data, making it appear more benign.

Conclusions:

• AnoDDPMs demonstrate significant potential for unsupervised OOD detection in digital pathology.
• The method shows flexibility with varying information bottlenecks and signal-to-noise ratios.
• While not matching fully supervised methods, AnoDDPMs represent a promising advancement for medical imaging anomaly detection.