High-performance Data Management for Whole Slide Image Analysis in Digital Pathology

Area of Science:

• Digital Pathology
• Computational Imaging
• Data Management

Background:

• Whole-slide imaging (WSI) in digital pathology generates massive datasets.
• Patch-level processing of WSIs creates significant input-output (I/O) bottlenecks.
• Efficient data access is crucial for timely analysis of gigapixel pathology images.

Purpose of the Study:

• To address the I/O challenges in digital pathology whole-slide image analysis.
• To implement and evaluate the Adaptable IO System version 2 (ADIOS2) for streamlined data management.
• To develop strategies for reducing data retrieval times in pathology workflows.

Main Methods:

• Developed a digital pathology-centric pipeline utilizing ADIOS2.
• Implemented strategies to optimize data access and retrieval.
• Conducted performance evaluations in both CPU-based and GPU-based scenarios.

Main Results:

• ADIOS2 achieved a two-fold speed-up in CPU-based image analysis compared to brute-force methods.
• In GPU-based deep learning scenarios, ADIOS2 demonstrated performance comparable to NVIDIA Magnum IO GPU Direct Storage (GDS).
• This represents an early application of ADIOS2 in the digital pathology domain.

Conclusions:

• ADIOS2 effectively mitigates I/O bottlenecks in gigapixel whole-slide image analysis.
• The system offers significant performance improvements for both CPU and GPU-based pathology workflows.
• The public release of the source code facilitates further research and adoption in digital pathology.