Scalable Whole Slide Image Reader Using Parallel GPU Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Whole slide images (WSIs) pose challenges due to their large size, requiring high computational overhead and time for processing, and existing technologies like OpenSlide are bottlenecks in deep learning applications, especially when dealing with large datasets for research and clinical purposes.
Innovation Solution
A scalable system utilizing parallel processing on both CPUs and GPUs with unbuffered tile reads, where each tile is allocated memory corresponding to its size, eliminating the need for buffering and reducing memory usage, allowing for efficient decoding and processing of WSIs without re-encoding, thereby improving disk performance and computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If existing technologies like OpenSlide are used to process whole slide images, then the images can be viewed and shared, but the processing speed is slow and computational overhead is high
Solution Approach 1:
The patent divides the whole slide image into multiple tiles and processes them in parallel using multiple CPU threads and GPU streams. Each tile is independently decoded and rendered, allowing concurrent processing that significantly improves throughput while reducing the computational burden on any single processing unit.
Solution Approach 2:
The patent replaces the traditional CPU-based decoding mechanism with GPU-based decoding using CUDA kernels. This substitution leverages the parallel processing capabilities of the GPU to accelerate the computationally intensive decoding operations, achieving substantial speedup over conventional CPU-based approaches.
2Productivity
If buffered reading is used to ensure data availability, then processing can be continuous, but memory usage increases significantly
Solution Approach 1:
The patent implements dynamic memory management where buffer sizes are adjusted based on actual processing needs rather than allocating fixed large buffers. The system dynamically allocates memory for tile batches and uses streaming techniques to process tiles as they become available, maintaining high throughput while minimizing peak memory usage.
Solution Approach 2:
The patent performs preliminary organization of tile data with metadata (position indicators, memory size indicators) before processing. This preliminary action allows the system to efficiently allocate memory only when needed and to process tiles in an optimized sequence without requiring large pre-allocated buffers.
3Measurement precision
If larger tissue patches are retrieved for deep learning analysis, then accuracy improves, but processing time and memory consumption increase
Solution Approach 1:
The patent segments large tissue patches into smaller tile units that can be processed in parallel. By dividing the large patch into multiple tiles and processing them concurrently on GPU streams, the system maintains the ability to analyze large regions for deep learning while reducing the time required through parallelization.
Solution Approach 2:
The patent adds the dimension of parallel processing by utilizing multiple GPU streams and CPU threads simultaneously. This allows the system to process multiple tiles of large tissue patches at the same time, effectively converting a sequential time-consuming operation into a parallel operation that achieves similar results much faster.
Data Source
AI summary
Described are embodiments of a scalable system for processing whole slide images (WSIs). The system may heavily utilize parallel processing on both central processing units (CPUs) and graphics processing units (GPUs). Images may be decoded on a GPU. Reads may be unbuffered with a low memory footprint. Compute Unified Device Architecture (CUD A) operations may be employed to eliminate many inefficiencies in traditional WSI processing libraries.