Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Tomasz Markiewicz1,2, Zaneta Swiderska-Chadaj1,,
Bartlomiej Grala2, Piotr Murawski3 ,Andrzej Kowalski3
1 Warsaw University of Technology, Dept. of Electrical Eng.
and
Military Institute of Medicine, 2Department of Pathomorphology, and 3Department of Information
Technology
Brain tumours
The meningiomas and oligodendrogliomas are the most frequent primary intracranial tumours. We can classify them into:
meningothelial (WHO I), atypical (WHO II), anaplastic (WHO III),
oligodendrogliomas (WHO II and III).
Based on this classification, we can h make a prognosis and choose an optimal therapy that correlates the tumor proliferation.
Introduction
Materials
Methods
Result
Conclusion
Specimens
The quantitative examination of histological tissues subject to immunostain tests is a basic method of recognizing a tumor, choosing optimal therapy and defining the prognostic indicators.
The value of this marker reflects the rate of tumor cell proliferation, and indicates the speed of tumour growth, as well as the degree of malignancy.
Introduction
Materials
Methods
Result
Conclusion
Stain and evaluation One of the most important
marker is the proliferation marker Ki-67/MIB-1, the one which is widely used to evaluate tumour.
The Ki67/MIB-1 marks the immunopositive cell nuclei with brown color whereas the other cell nuclei are marked with blue.
The quantitative evaluation of the specimen is required to establish an proliferation index.
Introduction
Materials
Methods
Result
Conclusion
Specimen, WSI and FOV
WSI FOV/ HOT-SPOT
Introduction
Materials
Methods
Result
Conclusion
Web-based platform Analysis of the whole slide images of immunohistochemically stained brain tumour specimens is demanding from both algorithmic and computational perspectives.
Introduction
Materials
Methods
Result
Conclusion
Proliferation For the detection of maximal proliferation regions of a brain tumour called a hot-spot, we use:
the mathematical morphology methods,
textural descriptions,
classification
penalty function
Introduction
Materials
Methods
Result
Conclusion
Algorithm
Introduction
Materials
Methods
Result
Conclusion
Size reduction Due to very large size of images (for example 100 000×80 000 pixels) in the contextual analysis of the specimen, it was necessary to reduce the resolution to enable direct examination and visualization.
1x 2x 4x 8x 16x
Introduction
Materials
Methods
Result
Conclusion
The aim of the study
The whole slide image, even in the reduced resolution, needs a dozen of minutes for full analysis. So, the acceleration of that process is necessary.
To complite this task, we propose the CPU-GPU processing scheme.
Introduction
Materials
Methods
Result
Conclusion
Materials 200 cases of meningiomas and oligodendrogliomas
subject to Ki-67/MIB-1 immunohistochemical staining were obtained from the archives of Department of Pathomorphology from the Military Institute of Medicine in Warsaw, Poland
Acquisition of the whole slide images was performed on the 3DHistech Pannoramic 250 Flash II scanner. The images were acquired under magnification 200x with a resolution 0.38 µm per pixels
Introduction
Materials
Methods
Result
Conclusion
Methods
The original sequential algorithm written in Matlab for the hot-spot detection was parallelized based on the Parallel Computing Toolbox for the division of computations between CPU and GPU.
The platform for computation tasks is composed with the dual processor unit of Intel Xeon EC-2650 2.60GHz (8 cores, 16 treads in each of them) supported by the two Tesla K40m GPU devices (2 800 cores, 12 GB).
Introduction
Materials
Methods
Result
Conclusion
GPU:List of Supported Functions
bwdist imabsdiff imgradientxy mat2gray
bwlabel imadjust imhist mean2
bwlookup imbothat imlincomb medfilt2
bwmorph imclose imnoise normxcorr2
corr2 imcomplement imopen padarray
edge imdilate imreconstruct radon
histeq imerode imregdemons rgb2gray
im2double imfill imresize rgb2ycbcr
im2int16 imfilter imrotate std2
im2single imgaussfilt imshow stdfilt
im2uint8 imgaussfilt3 imtophat stretchlim
im2uint16 imgradient iradon ycbcr2rgb
Introduction
Materials
Methods
Result
Conclusion
Methods Algorithm
O Algorithm
A Algorithm
B Algorithm
C Algorithm
D Algorithm
E
Original algorithm without
multithreading
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Full multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Send whole image to GPU, out
of memory problem
Tiled WSI computing, multitreadi
ng of all FOVs, each
send to GPU
independently
Introduction
Materials
Methods
Result
Conclusion
Methods Algorithm
O Algorithm
A Algorithm
B
Original algorithm without
multithreading
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Introduction
Materials
Methods
Result
Conclusion
Methods Algorithm
O Algorithm
A Algorithm
B Algorithm
C
Original algorithm without
multithreading
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Full multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Introduction
Materials
Methods
Result
Conclusion
Methods Algorithm
O Algorithm
A Algorithm
B Algorithm
C Algorithm
D
Original algorithm without
multithreading
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Full multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Send whole image to GPU, out
of memory problem
Introduction
Materials
Methods
Result
Conclusion
Methods Algorithm
O Algorithm
A Algorithm
B Algorithm
C Algorithm
D
Original algorithm without
multithreading
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Full multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Send whole image to
GPU, out of
memory problem
Introduction
Materials
Methods
Result
Conclusion
Algorithms
Algorithm A
Algorithm B
Algorithm C
Algorithm E
Partial multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Multithreading
analysis of all FOV on
CPU
Full multithrea
ding on CPU, image analysis as a
tape with 1000 pixel width and whole WSI length (10 000 pixels)
Tiled WSI computing, multitreadi
ng of all FOVs, each
send to GPU
independently
Results – computational time Algorithm O Algorithm A Algorithm B Algorithm C Algorithm E
646 201 242 280 89
Introduction
Materials
Methods
Result
Conclusion
Algorithm E
Methods
Result
Conclusion
FOV vs Calculation time
Introduction
Materials
Methods
Result
Conclusion
Memory Usage
Introduction
Materials
Methods
Result
Conclusion
Conclusions
When the sequential analysis needs a dozen to two dozens of minutes, then:
only a parallelization on CPU gives a triple time reduction approximately,
the proper division of computation tasks between CPU and GPU resulted in the decrease of computational time by seven to ten times on average.
Introduction
Materials
Methods
Result
Conclusion
646- 89 s
Seven times shorter calculation time
Conclusions The proper allocation of computational tasks between CPU and GPU gives possibilities of high acceleration of virtual slide analysis in pathomorphological practice. The type, complexity, and support of various data transformation is crucial to planning and data management.
Introduction
Materials
Methods
Result
Conclusion
Acknowledgement: This study was supported by the National Centre for Research and Development, Poland (grant PBS2/A9/21/2013).
Thank you for your attention