Deep-learning triage of 3D pathology datasets for comprehensive and efficient pathologist assessments

Area of Science:

• Computational pathology
• Digital pathology
• Artificial intelligence in medicine

Background:

• Standard histopathology (2D) undersamples heterogeneous tissues, with each section representing <1% of total tissue volume.
• 3D pathology techniques, like open-top light-sheet microscopy (OTLS), offer comprehensive, high-resolution imaging of large specimens.
• Manual review of massive 3D datasets is infeasible for routine clinical pathology.

Purpose of the Study:

• To develop a deep learning framework (CARP3D) for efficient pathologist evaluation of 3D pathology data.
• To enable accelerated clinical adoption of 3D pathology by focusing pathologist attention on high-risk areas.
• To improve disease detection and optimize pathologist workload in clinical settings.

Main Methods:

• CARP3D, a deep learning triage framework, was developed to identify high-risk 2D cross-sections within 3D pathology datasets.
• The framework assigns risk scores to 2D levels by incorporating contextual information from neighboring depth levels.
• Performance was evaluated using two use cases: prostate cancer biopsy risk stratification and Barrett's esophagus dysplasia/cancer screening.

Main Results:

• CARP3D successfully identifies high-risk 2D cross-sections within large 3D pathology datasets.
• Leveraging neighboring depth context improved prediction accuracy compared to models using isolated 2D levels.
• AI-triaged 3D pathology demonstrated potential for improved high-risk disease detection and optimized pathologist workload.

Conclusions:

• CARP3D facilitates time-efficient pathologist evaluation of large 3D pathology datasets.
• This approach enhances the potential of 3D pathology for improved disease detection and clinical decision-making.
• CARP3D represents a viable pathway for integrating advanced 3D pathology into routine clinical practice.