Accelerated Image Processing Using Graphical Processing Units

Document Type


Degree Name

Master of Science (MS)


Computer Science and Info Sys

Date of Award

Summer 2012


Histopathological examination of the cancer tissues plays an important role in diagnosis and prognosis of critical diseases. The microvessel density for instance, is an indicator for the prognosis of the cancer. In spite of significant advancements in high resolution image scanning, the analysis of the images is still done manually which takes considerable amount of time. With the advent of graphical processing units (GPUs) in general purpose applications, it has been shown that the time taken to perform the computational tasks can be reduced significantly. Regardless of widely available GPU algorithms, many algorithms and frameworks are to be redesigned by computational researchers to serve their purposes. The GPUs are emerged as general purpose computation on graphical processing units (GPGPU) that can be used to perform computation in applications handled by central processing unit (CPU). This study proposes a platform on which analysis of high resolution images are done by graphical processing units in relatively shorter time than CPU. The image processing methods implemented are neighborhood search, color classification of pixels, convolution within a 3x3 window, maximum and minimum pixel value with in a 3x3 window and a feature to identify microvessels in the tumor. The image processing methods implemented on GPU produced faster results when compared to general serial CPU computation. The proposed image processing methods were tested on sample of 585 images collected from two high resolution virtual slides. The image processing methods are implemented on GPU and attained significant speedup when compared to CPU versions. In this study, we show that processing of high resolution images is successfully adapted into GPU based algorithms and introduce a feature for identifying the microvessels which helps pathologists investigate the prognosis of the cancer. Image processing methods implemented showed significant speedups can be obtained, such as a speedup of approximately 40x in neighborhood search, 50x in convolution, more than 200x in color classification of pixels, 30x in maximum pixel value replacement, and 50x in a feature extraction operation for microvessel identification.


Mutlu Mete

Subject Categories

Computer Sciences | Physical Sciences and Mathematics