High-throughput Analysis of Large Microscopy Image Datasets on CPU-GPU Cluster Platforms.

Area of Science:

• Computational pathology
• High-performance computing
• Medical image analysis

Background:

• Pathology image analysis is crucial for disease morphology studies but is limited by computational resource demands.
• Large-scale analysis of high-resolution pathology images requires efficient, scalable computational pipelines.

Purpose of the Study:

• To develop and evaluate a parallel image analysis pipeline for high-throughput computation of large pathology image datasets.
• To optimize performance on distributed CPU-GPU platforms for cancer image analysis.

Main Methods:

• Implemented a hierarchical data processing pipeline with coarse-grain stages and fine-grain operations.
• Utilized performance-aware scheduling, architecture-aware placement, data locality, prefetching, and asynchronous data copy.
• Employed a hybrid CPU-GPU computing approach for parallel processing.

Main Results:

• Cooperative CPU-GPU utilization improved performance by up to 1.6x compared to GPU-only versions.
• Fine-grain operation scheduling offered superior runtime optimization opportunities over monolithic implementations.
• Processed 1.8TB of image data (36,848 tiles) in under 4 minutes on a 100-node cluster.

Conclusions:

• The proposed parallel pipeline effectively addresses computational bottlenecks in large-scale pathology image analysis.
• Hybrid CPU-GPU computing and fine-grain parallelism are key to achieving high throughput and efficiency.
• This approach enables faster and more comprehensive investigations of disease morphology in large image datasets.