Developing image analysis pipelines of whole-slide images: Pre- and post-processing.

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Deep learning advances digital pathology beyond basic digitization and telemedicine.
• Integration of deep learning into routine laboratory workflows is imminent and may revolutionize pathology.
• Current focus on computational methods overlooks essential workflow integration steps.

Purpose of the Study:

• To review methods for deep learning in digital pathology, focusing on image analysis.
• To discuss the challenges and unique issues in analyzing very large images, such as whole-slide images (WSIs).
• To highlight the necessity of pre- and post-processing for deep learning model interpretation and prediction.

Main Methods:

• Review of image analysis techniques for deep learning in digital pathology.
• Discussion of pre-processing steps: artifact detection, color normalization, subsampling, tiling.
• Examination of post-processing steps: removal of errant predictions, interpretation of results.
• Addressing challenges posed by large image file sizes (gigabytes) requiring tiling for whole-slide image (WSI) analysis.

Main Results:

• Deep learning integration requires significant pre- and post-processing beyond model training.
• Image tiling is a common necessity due to the large file sizes of whole-slide images (WSIs).
• Effective analysis necessitates robust methods for artifact detection, normalization, and prediction refinement.

Conclusions:

• Successful implementation of deep learning in digital pathology hinges on comprehensive image processing pipelines.
• Addressing the unique challenges of large-format image analysis is crucial for clinical adoption.
• Deep learning holds potential to reduce processing time and improve anomaly detection in pathology.