Medical Image Processing and GPU Programming - Research Papers

Segmentation of Colon Tissue Samples using Graphics Accelerators

Nowadays microscopic analysis of tissue samples is done more and more by using digital imagery and special immunodiagnostic software. The main aim of this work is to show, how GPGPUs can facilitate certain type of image processing methods. The software developed by the Biotech Group is used to detect special tissue parts on HE (hematoxilin eosin) stained colon tissue sample images. Since pathologists are working with large number of high resolution images - thus require significant storage space -, one feasible way to achieve reasonable processing time is the usage of GPGPUs. The CUDA software development kit was used to develop processing algorithms to NVIDIA type GPUs.

Healty colon tissue Malicious colon tissue
Healthy colon tissue Malicious colon tissue (CRCa)

Detection of cell nuclei

Several segmentation procedures are based on the segmentation of the image and a lot of them need the number and the locations of the cells. The aim of our research is developing a new data parallel algorithm that can be implemented even in a GPGPU environment and that is capable of counting hematoxylin eosin (HE) stained cell nuclei and of identifying their exact locations and sizes (using a variation of the region growing method). The new method has three levels of parallelization:

  o Parallelization of the region growing method itself to use 1 thread for processing of each contour points.
  o Starting more than one region growings in the GPGPU at the same time to fully utilize the processing power.
  o Using multiple GPGPUs based on the split-and-merge method.

As our test results affirmed, the GPU usage was able to speed up the image processing tasks significantly. It can be stated that (even thought the smaller images can be processed faster with the graphics unit as well) the real benefits of the GPU come to the front at higher resolution.
Cell nuclei

Detection of epithelial cells

Epithelial tissues line the cavities and surfaces of structures throughout the body. They also form many glands. In HE stained colon tissue samples, epithelial cells appear around the glands and at the edge of the whole sample (surface epithelium). After the cell nuclei detection, we have to determine that an appropriate nucleus belongs to an epithelial cell or not. The method developed by us is based on the idea that that the epithelial cells have some particular attributes compared to the other cells:

  o Density of the epithelial cell nuclei groups differs from the average density of the other cells of the tissue sample:
     • High density in one direction.
     • Relatively small density orthogonally.
     • Average density in the other orthogonal direction.
  o Epithelial cells usually form a chain.
Cell nuclei

Detection of glands

The gland is formed by a chain of epithelial cells. The colon adenocarcinoma causes changes in glandular structures of colon tissues, therefore for grading we can use the gland locations. There are several, already existing promising procedures for this segmentation, these make use of the fact that glands are characterized by their luminal areas surrounded by the epithelial cells. Our solution based on the identified epithelial cell nuclei. We have to decide that a given epithelial cell nucleus is belongs to a gland or to the surface epithelium.

This research is still going on, we have only some preliminary results.
Cell nuclei

Our publications related to this topic

Title Parallel Biomedical Image Processing with GPGPUs in Cancer Research
Authors Reményi, A., Szénási, S., Bándi I., Vámossy Z., Valcz G., Bogdanov P., Sergyán S., Kozlovszky M.
Appears in Reményi, A., Szénási, S., Bándi I., Vámossy Z., Valcz G., Bogdanov P., Sergyán S., Kozlovszky M., "Parallel Biomedical Image Processing with GPGPUs in Cancer Research", 3rd IEEE International Symposium on Logistics and Industrial Informatics, Budapest, 25-27 Aug. 2011, pp. 245-248, ISBN 978-1-4577-1840-3
Abstract The main aim of this work is to show, how GPGPUs can facilitate certain type of image processing methods. The software used in this paper is used to detect special tissue part, the nuclei on (HE - hematoxilin eosin) stained colon tissue sample images. Since pathologists are working with large number of high resolution images - thus require significant storage space -, one feasible way to achieve reasonable processing time is the usage of GPGPUs. The CUDA software development kit was used to develop processing algorithms to NVIDIA type GPUs. Our work focuses on how to achieve better performance with coalesced global memory access when working with three-channel RGB tissue images, and how to use the on-die shared memory efficiently.
Download Download full paper as pdf or from IEEE Xplore.

Title GPGPU-based data parallel region growing algorithm for cell nuclei detection
Authors Szénási, S., Vámossy, Z., Kozlovszky, M.
Appears in Szénási, S., Vámossy, Z., Kozlovszky, M., "GPGPU-based data parallel region growing algorithm for cell nuclei detection", 12th IEEE International Symposium on Computational Intelligence and Informatics (CINTI), Budapest, 21-22 Nov. 2011, pp.493-499, ISBN 978-1-4577-0044-6
Abstract Nowadays microscopic analysis of tissue samples is done more and more by using digital imagery and special immunodiagnostic software. These are typically specific applications developed for one distinct field, but some subroutines are commonly repeated, for example several applications contain steps that can detect cell nuclei in a sample image. The aim of our research is developing a new data parallel algorithm that can be implemented even in a GPGPU environment and that is capable of counting hematoxylin eosin (HE) stained cell nuclei and of identifying their exact locations and sizes (using a variation of the region growing method). Our presentation contains the detailed description of the algorithm, the peculiarity of the CUDA implementation, and the evaluation of the created application (regarding its accuracy and the decrease in the execution time).
Download Download full paper as pdf or from IEEE Xplore.

Title Evaluation and comparison of cell nuclei detection algorithms
Authors Szénási, S., Vámossy, Z., Kozlovszky, M.
Appears in Szénási, S., Vámossy, Z., Kozlovszky, M., "Evaluation and comparison of cell nuclei detection algorithms", 16th IEEE International Conference onIntelligent Engineering Systems (INES), Lisbon, 13-15 June 2012, pp.469-475, ISBN 978-1-4673-2694-0
Abstract The processing of microscopic tissue images and especially the detection of cell nuclei is nowadays done more and more using digital imagery and special immunodiagnostic software products. Since several methods (and applications) were developed for the same purpose, it is important to have a measuring number to determine which one is more efficient than the others. The purpose of the article is to develop a generally usable measurement number that is based on the “gold standard” tests used in the field of medicine and that can be used to perform an evaluation using any of image segmentation algorithms. Since interpreting the results themselves can be a pretty time consuming task, the article also contains a recommendation for the efficient implementation and a simple example to compare three algorithms used for cell nuclei detection.
Download Download full paper as pdf or from IEEE Xplore.

Title Preparing initial population of genetic algorithm for region growing parameter optimization
Authors Szénási, S., Vámossy, Z., Kozlovszky, M.
Appears in Szénási, S., Vámossy, Z., Kozlovszky, M., "Preparing initial population of genetic algorithm for region growing parameter optimization", 4th IEEE International Symposium on Logistics and Industrial Informatics (LINDI), Smolenice, 5-7 Sept. 2012, pp.47-54, ISBN 978-1-4673-4520-0
Abstract The processing of microscopic tissue images is nowadays done more and more using special immunodiagnostic-evaluation software products. Often to evaluate the samples, the first step is determining the number and location of cell nuclei. To do this, one of the most promising methods is the region growing, but this algorithm is very sensitive to the appropriate setting of different parameters. Due to the large number of parameters and due to the big set of possible values setting those parameters manually is a quite hard task, so we developed a genetic algorithm to optimize these values. The first step of the development is the statistical analysis of the parameters, and the determination of the important features, to extract valuable information for a to-be-implemented genetic algorithm that will perform the optimization.
Download Download full paper as pdf or from IEEE Xplore.

Title Parameter assisted HE colored tissue image classification
Authors Kozlovszky, M., Hegedűs, K., Szénási, S., Kiszler, G., Wichmann, B., Bandi, I., Eigner, Gy., Sas, P. I., Kovács, L., Garaguly, Z., Kiss, G., Valcz, G., Molnár, B.
Appears in Kozlovszky, M., Hegedűs, K., Szénási, S., Kiszler, G., Wichmann, B., Bandi, I., Eigner, Gy., Sas, P. I., Kovács, L., Garaguly, Z., Kiss, G., Valcz, G., Molnár, B., "Parameter assisted HE colored tissue image classification", 17th International Conference on Intelligent Engineering Systems (INES 2013), Costa Rica, 19-21 Jun. 2013, pp.203-207, ISBN 978-1-4799-0830-1
Abstract The aim of our work was to design and implement a software solution, which supports quantitative histological analysis of hematoxilin eozin (HE) stained colon tissue samples, identify tissue structures – nuclei, glands and epithelium – using image processing methods. Furthermore, based on the result of the histological segmentation, it gives a suggestion for the negative or malignant status of the samples automatically. In this paper we describe the algorithm which builds up mainly by two software components: MorphCheck -our software framework-, which is capable to make effective, morphometric evaluation of high resolution digital tissue images and a modified WND-CHARM (Weighted Neighbor Distance Using Compound Hierarchy of Algorithms Representing Morphology), which is a multi-purpose image classifier. The image classification was performed mainly based on 75+15 pre-defined colon tissue specific parameters, which were measured by MorphCheck, and other 2873 in-built generic image parameters, which were measured by WND-CHARM. We appended WND-CHARM’s learning and classification capabilities with our colon tissue specific parameters and with this act we have increased its classification accuracy significantly on HE stained colon tissue sample images.
Download Download full paper as pdf or from IEEE Xplore.

Title Colon cancer diagnosis on digital tissue images
Authors Kerekes, Z., Tóth, Z., Szénási S., Vámossy Z., Sergyán Sz.
Appears in Kerekes, Z., Tóth, Z., Szénási S., Vámossy Z., Sergyán Sz., "Colon cancer diagnosis on digital tissue images", 2013 IEEE 9th International Conference on Computational Cybernetics (ICCC2013), Tihany, 8-10. July 2013, pp.159-163, ISBN 978-1-4799-0060-2
Abstract The purpose of this project is to develop a software which can be an aid for difficult colon cancer diagnosis and using this system the patients can be helped with an early diagnosis. The aim can be achieved with processing and analysing microscopic tissue images. This paper contains the basic knowledges related to the project and the description of the developed system. The implemented algorithms determines the locations and features of glands and save these information for the subsequent diagnosis. One of the most important algorithm in this project is the Color Structure Code, which performs a color based segmentation and the output is the starting point of the further process.
Download Download full paper as pdf or from IEEE Xplore.

Title Medical Image Segmentation with Split-and-Merge Method
Authors Szénási, S.
Appears in Szénási, S., "Medical Image Segmentation with Split-and-Merge Method", 5th IEEE International Symposium on Logistics and Industrial Informatics (LINDI 2013), Wildau, 5-7 Sept. 2013, pp.137-140, ISBN 978-1-4799-1258-2/13
Abstract The processing of microscopic tissue images and especially the detection of cell nuclei is nowadays done more and more using digital imagery and special immunodiagnostic software products. One of the most promising methods is region growing but it is quite memory intensive. The size of high-resolution tissue images can easily reach the order of a hundred megabytes therefore the memory requirement for the region growing is more than one gigabyte. To provide the execution in low-end clients we have to split the whole image into smaller tiles and after the processing of each individual tiles we have to merge the results.
Download Download full paper as pdf.

Title Implementation of a Distributed Genetic Algorithm for Parameter Optimization in a Cell Nuclei Detection Project
Authors Szénási, S., Vámossy, Z.
Appears in Szénási, S., Vámossy, Z., "Implementation of a Distributed Genetic Algorithm for Parameter Optimization in a Cell Nuclei Detection Project", Acta Polytechnica Hungarica, 2013, Vol.10, No.4, pp.89-86, ISSN 1785-8860
Abstract The processing of microscopic tissue images and especially the detection of cell nuclei is nowadays done more and more using digital imagery and special immunodiagnostic software products. One of the most promising image segmentation method s is region growing, but this algorithm is very sensitive to the appropriate setting of different parameters , and the long runtime due to its high computing demand reduces its practical usability. As a result of our research , we managed to develop a data - p arallel region growing algorithm that is two or three times faster than the original sequential version . The paper summarizes o ur results : the development of an evolution - based algorithm that was used to successfully determine a set of parameters that coul d be used to achieve significantly better accuracy than the already existing parameters.
Download Download full paper as pdf.

Title Evolutionary Algorithm for Optimizing Parameters of GPGPU-based Image Segmentation
Authors Szénási, S., Vámossy, Z.
Appears in Szénási, S., Vámossy, Z., "Evolutionary Algorithm for Optimizing Parameters of GPGPU-based Image Segmentation", Acta Polytechnica Hungarica, 2013, Vol.10, No.5, pp.7-28, ISSN 1785-8860
Abstract The use of digital microscopy allows diagnosis through automated quantitative and qualitative analysis of the digital images. Often to evaluate the samples, the first step is determining the number and location of cell nuclei. For this purpose, we have developed a GPGPU based data-parallel region growing algorithm that is equally as accurate as the already existing sequential versions, but its speed is two or three times faster (implementing in CUDA environment), but this algorithm is very sensitive to the appropriate setting of different parameters. Due to the large number of parameters and due to the big set of possible values setting those parameters manually is a quite hard task, so we have developed a genetic algorithm to optimize these values. Our evolution-based algorithm that is described in this paper was used to successfully determine a set of parameters that compared to the results with the previously known best set of parameters means a significantly improvement.
Download Download full paper as pdf.

Title Genetic Algorithm for Parameter Optimization of Image Segmentation Algorithm
Authors Szénási, S.
Appears in Szénási, S., "Genetic Algorithm for Parameter Optimization of Image Segmentation Algorithm", 14th IEEE International Symposium on Computational Intelligence and Informatics (CINTI 2013), Budapest, 19-21 Nov. 2013, pp.351-354, ISBN 978-1-4799-0197-5/13
Abstract In the current practice of medicine, histopathological examinations are some of the most important tools for clinical diagnoses of a large group of diseases. To help pathologists and to reduce the subjectivity level, it has been proposed that computer-aided procedures be used to provide objective results. The first step of these procedures is the segmentation of the tissue image. In our research, we try to detect nuclei, glands and surface epithelium in Haematoxylin and Eosin (HE) stained colon tissue samples. This paper focuses on the identification of epithelial cell nuclei.
Download Download full paper as pdf.

Title Distributed Implementations of Cell Nuclei Detection Algorithm
Authors Szénási, S.
Appears in Szénási, S., "Distributed Implementations of Cell Nuclei Detection Algorithm", 1st International Conference on Image Processing and Pattern Recognition (IPPR '13), Budapest, 10-12 Dec. 2013, pp.105-109, ISBN 978-960-474-350-6
Abstract One of the most promising methods for cell nuclei detection in colon tissue images is region growing, but it has several limitations. The process is slow to the extent that practical use seems almost impossible since the segmentation of large images that contain many nuclei may require up to 40-60 minutes to complete. However, the method is very promising, it offers very good accuracy; therefore, it is definitely worth dealing with this drawback. However, we have tried to speed up the process based on distributed execution using some novel techniques: a naive implementation, a compatible synchronized version, and an implementation based on the split-and merge technique. This paper presents the “compatible synchronized implementation” in detail.
Download Download full paper as pdf.

Title Distributed Region Growing Algorithm for Medical Image Segmentation
Authors Szénási, S.
Appears in Szénási, S., "Distributed Region Growing Algorithm for Medical Image Segmentation", International Journal of Circuits, Systems and Signal Processing, 2014, Vol. 8, No. 1, pp.173-181, ISSN 1998-4464
Abstract Signal processing plays an important role in the work of pathologists; it is especially true for image processing software products. High-resolution digital images have taken over the role of traditional tissue slides on a glass plate. In addition to the direct effects of this advancement (sharing images, remote access, etc.), a new option appeared: the possibility of using image processing software for automatic (or semi-automatic) diagnostics. One of the most important tasks in this procedure is the segmentation of the tissue images; we have to identify the main components (in the case of colon tissue samples, these are the cell nuclei, glands and surface epithelium). There are several traditional image segmentation methods for this purpose, but none of them provides both acceptable accuracy and runtime. This paper presents a distributed region growing method implemented on CPUs and GPGPUs.
Download Download full paper as pdf.

Title Segmentation of Colon Tissue Sample Images Using Multiple Graphics Accelerators
Authors Szénási, S.
Appears in S. Szénási, "Segmentation of colon tissue sample images using multiple graphics accelerators", Computers in Biology and Medicine (2014), Vol. 51, pp. 93-103, DOI 10.1016/j.compbiomed.2014.05.002
Abstract Nowadays, processing medical images is increasingly done through using digital imagery and custom software solutions. The distributed algorithm presented in this paper is used to detect special tissue parts, the nuclei on haematoxylin and eosin stained colon tissue sample images. The main aim of this work is the development of a new data-parallel region growing algorithm that can be implemented even in an environment using multiple video accelerators. This new method has three levels of parallelism: a) the parallel region growing itself b) starting more region growing in the device c) using more than one accelerator. We use the split-and-merge technique based on our already existing data-parallel cell nuclei segmentation algorithm extended with a fast, backtracking-based, non-overlapping cell filter method. This extension does not cause significant degradation of the accuracy; the results are practically the same as those of the original sequential region growing method. However, as expected, using more devices usually means less time is needed to process the tissue image; in the case of the configuration of one central processing unit and two graphics cards, the average speed-up is about 4–6X. The implemented algorithm has the additional advantage of efficiently processing very large images with high memory requirements.
Download Download full paper as pdf (Pre-print) or from ScienceDirect (Published article).

Title Optimizing General Purpose Computations Using Kepler Based Graphics Accelerators
Authors Szénási, S.
Appears in Szénási, S., "Optimizing General Purpose Computations Using Kepler Based Graphics Accelerators", International Masaryk Conference for PhD students and young researchers (MMK2014), Hradec Králové, 15-19 Dec. 2014, pp. 3354-3360, ISBN 978-80-87952-07-8
Abstract The programming of GPUs (Graphics Processing Units) is ready for practical applications; the largest industry players (including research centres, financial and analyst corporations) have already announced that they use these new devices for high computing applications. There are several well-known areas, like image processing, simulations and obviously 3D graphics, where we can use these devices very efficiently. In this paper, we would like to show, that beyond these well-known topics, GPU programming is able to speed-up more general purpose applications. The key is the data parallel nature of the algorithm, and the minimization of data transfers between CPU and GPU.
Download Download full paper as pdf.

Title Solving Multiple Quartic Equations on the GPU using Ferrari's Method
Authors Szénási, S. and Tóth, Á.
Appears in Szénási, S., Tóth, Á., "Solving Multiple Quartic Equations on the GPU using Ferrari's Method", IEEE 13th International Symposium on Applied Machine Intelligence and Informatics, Herlany, 22-24 Jan. 2015, pp. 333-337, ISBN 978-1-4799-8221-9
Abstract As known, quartics are the highest degree polynomials which can be solved analytically in general by the methods of radicals. There are several problems based on not only one but more equations independently, in case of simulations, the number of equations can be very high. For this reason, it is worth examining the runtime of the solver algorithms implemented for multi-core systems, especially graphics accelerators. In this paper, we discuss the runtime and numerical stability of the Ferrari's method using GPUs. It is worth to port an application to the graphics card, if the number of calculations is relatively high and the number and volume of memory accesses is relatively small. Based on the results, it is clear, that running multiple equation solvers based on the given method is clearly meets these conditions.
Download Download full paper as pdf.

Acknowledgement

Óbuda University 3DHISTECH Ltd.
Obuda University 3DHISTECH Ltd.