Parallel and Efficient Sensitivity Analysis of Microscopy Image Segmentation Workflows in Hybrid Systems.

Area of Science:

• Computational imaging
• High-performance computing
• Algorithm optimization

Background:

• Sensitivity analysis (SA) is crucial for evaluating image analysis algorithms but is computationally intensive.
• Large datasets of high-resolution images require efficient SA methods for practical application.
• Current SA methods often face performance bottlenecks due to repeated computations.

Purpose of the Study:

• To introduce and evaluate strategies for accelerating SA of image segmentation and classification algorithms.
• To enable efficient SA on large-scale datasets using distributed hybrid systems.
• To quantify the performance gains from runtime optimizations and computation reuse.

Main Methods:

• Implemented runtime optimizations targeting distributed hybrid systems (Intel Phi and CPUs).
• Employed smart task assignment strategies for cooperative execution on hybrid nodes.
• Utilized multi-level computation reuse to avoid redundant calculations.
• Evaluated the approach on a cancer image analysis workflow using a 256-node cluster.

Main Results:

• Achieved over 90% parallel efficiency on 256 nodes.
• Gained an additional 2x speedup through cooperative CPU and Phi execution.
• Obtained up to 2.46x additional speedup using multi-level computation reuse.
• Demonstrated significant performance improvements for SA in large-scale studies.

Conclusions:

• The proposed optimizations significantly accelerate SA for image analysis algorithms.
• Efficient SA is now feasible for large-scale studies, particularly in medical imaging.
• Runtime optimizations and computation reuse are key to overcoming SA computational demands.