Computationally Efficient Pathologic Image Analysis: Use of GPUs for Classification of Stromal Density
Olcay Sertel MSc; The Ohio State University; Antonio Ruiz MSc; University of Malaga; Umit Catalyurek PhD; The Ohio State University; Manuel Ujaldon PhD; University of Malaga; Joel Saltz MD, PhD; The Ohio State University; Metin Gurcan PhD; The Ohio State University;
Content:
A goal of image analysis for microscopic pathology images is to investigate different types of measurable features through quantitative analysis that would lead to more objective, accurate and reproducible diagnostic clues. However, digitized pathological images are quite large, with sizes typically up to 80 k x 120 k, disc sizes around 50 GB. This limits the range of applications to be developed for clinical practice, especially those that require sophisticated image processing steps. Therefore, we are investigating novel hardware architectures to reduce the computational burden, which will allow quantitative image analysis to be applicable in clinical practice.
Technology:
A graphics processing unit (GPU) is a dedicated graphics rendering device for a PC or a workstation used primarily for 3D visualization applications. Although initially devoted to the gaming industry, GPUs have turned into a feasible platform for a wide spectrum of general-purpose applications due to their remarkable computational power, extraordinary bandwidth and flexible programmability. As part of a computer-assisted prognosis system, we are developing an image analysis system to classify the neuroblastoma images into stroma-rich and stroma-poor regions using the GPU.
Design:
A domestic GPU exceeds 500 GFLOPS of computational power nowadays, but its actual performance highly depends on how well it is programmed to exploit parallelism and memory bandwidth. We analyzed the trade-offs of processing each step of our image analysis algorithm on CPU and GPU by converting the Matlab code to C++ for CPU and then implementing the same code on GPU. Finally, we propose a parallel bi-processor platform combining the processing capabilities of CPU and GPU for a more efficient pathologic image analysis, where each processor is responsible for certain stages of the image analysis algorithm concerning the efficiency, workload balance and parallelism.
Results:
We tested the classification accuracy of our approach on a training set of 900 image tiles, 450 from both classes (i.e. stroma-rich and stroma-poor), that were randomly cropped from 20 different whole slide images. We applied leave-one-out and 10-fold cross validation and achieved remarkable classification accuracies of 99.33% and 98.4%, respectively. Moreover, the processing time of around 850 msec. for an individual test image tile on the CPU decreased by a factor of 45 times by introducing the bi-processor platform. This platform benefits from the best of CPU and GPU, thus complementing their features and preventing bottlenecks.
Conclusion:
By using the combined CPU/GPU processing platform, we achieved to decrease the run time of our pathologic image analysis application significantly. We envision that GPUs will continue to play an increasing role in processing large imaging data such as pathologic images.
