Revisiting Reconstruction-Based Anomaly Detection for Whole-Slide Image.

Area of Science:

• Computational pathology
• Medical image analysis
• Machine learning

Background:

• Whole Slide Images (WSIs) are crucial in computational pathology (CPath), but high-quality annotations are a bottleneck.
• Task-aware unsupervised anomaly detection models trained on normal data offer a promising alternative for tasks like cancer detection.
• Existing anomaly detection models often perform suboptimally on WSIs due to domain discrepancies.

Purpose of the Study:

• To analyze and adapt feature reconstruction-based anomaly detection models for WSIs.
• To propose an Explicit Conditional Reconstruction framework (ECR4AD) for enhanced performance in CPath.
• To evaluate ECR4AD's effectiveness in tile-level task-aware unsupervised anomaly detection.

Main Methods:

• Analysis of conditional reconstruction designs in anomaly detection.
• Refinement of existing models to better suit WSI characteristics.
• Development and implementation of the Explicit Conditional Reconstruction framework (ECR4AD).

Main Results:

• ECR4AD consistently achieved substantial improvements in Area Under the Receiver Operating Characteristic curve (AUROC) across four diverse datasets.
• The framework demonstrated effectiveness in breast and prostate cancer metastasis detection and prostate cancer Gleason grading.
• The proposed method shows superior performance for tile-level task-aware unsupervised anomaly detection in CPath.

Conclusions:

• ECR4AD offers a significant advancement for anomaly detection in WSIs within computational pathology.
• The framework effectively addresses the challenges of WSI analysis and annotation bottlenecks.
• ECR4AD provides a robust solution for task-aware unsupervised anomaly detection, enhancing diagnostic capabilities.