Τομέας Επικοινωνιών και Επεξεργασίας Σήματος
Library of the School of Science

2016-05-13

2016

Κατσιγιάννης Στάμος

Μαρούλης Δημήτρης,Θεοδωρίδης Σέργιος,Σαγκριώτης Εμμανουήλ

Advanced GPU-based methods for analysis and processing of biomedical images and general purpose video

English

Ανάπτυξη προηγμένων μεθόδων ανάλυσης και επεξεργασίας βιοϊατρικών εικόνων και βίντεο γενικού σκοπού με χρήση επεξεργαστών γραφικών

This thesis introduces a novel feature extraction method for ultrasound thyroid
texture classification, a novel method for complementary DNA (cDNA) microarray
image segmentation in order to assist the gene expression quantification
analysis, a GPU-based software tool for cDNA microarray image analysis and a
novel method for high quality real-time video compression on the GPU. GPU
approaches have been proposed in this thesis for all the developed methods in
order to significantly reduce the computational times needed and harness the
usually underutilized computational power of modern GPUs.
The first major contribution of this thesis is a novel feature extraction
method for ultrasound thyroid texture classification. The proposed method
relies on statistical feature extraction from ultrasound thyroid images in
order to assist the task of classification to healthy or nodular thyroid
tissue. Images are first decomposed using the Contourlet Transform and then a
set of statistical features is computed for the contourlet subbands along a
feature selection algorithm for selecting the most significant features and for
reducing the size of the feature vector. The classification accuracy achieved
using the proposed feature extraction method was evaluated on real ultrasound
thyroid images using Support Vector Machines (SVM). The experimental results
show that the proposed scheme performs better than state-of-the-art
alternatives, constituting a viable solution for the task of ultrasound thyroid
texture classification.
The second major contribution of this thesis is a novel method for
complementary DNA (cDNA) microarray image segmentation. The proposed algorithm
takes as input a cDNA microarray image and the pre-computed gridding results
and segments the image into background and cDNA spot regions. After a
pre-processing denoising step, initial seed pixels are automatically selected
for the background and spot regions and the final segmentation is computed
through a region growing scheme. Evaluation of the results on real cDNA
microarray images through visual inspection (due to the lack of ground-truth
information) shows that the proposed approach provides quality segmentation,
while statistical evaluation on synthetic cDNA microarray images shows that the
proposed approach outperforms state-of-the-art alternatives for cDNA microarray
image segmentation.
The third major contribution of this thesis is a software tool for cDNA
microarray image gridding and segmentation that offers an easy-to-use graphical
user interface that allows the analysis of the input images by a simple click
of a button. Computations are all performed on the GPU in order to offer
reduced computational times. The proposed software incorporates a proposed GPU
implementation of a very efficient cDNA microarray image gridding algorithm
that utilizes a genetic algorithm approach for performing the gridding of cDNA
microarray images. The proposed approach achieves significantly lower
computational times than the CPU approach and allows the use of the gridding
algorithm in a practical working scenario. In addition to the gridding
algorithm, the proposed software incorporates a GPU implementation of the
proposed cDNA microarray image segmentation algorithm. The proposed software
offers increased performance compared to state-of-the-art methods as it employs
advantageous algorithms that achieve enhanced results.
The final major contribution of this thesis is a novel method for high quality
real-time video compression on the GPU. The proposed algorithm introduces the
use of the Contourlet Transform in combination with various methods for lossy
and lossless compression for real-time video compression on the GPU. The
proposed approach achieves higher visual quality compared to other methods at
very low bitrates and does not suffer from blocking artifacts as in the case of
DCT-based compression algorithms. The most computationally intensive steps are
computed on the GPU, thus reducing the load of the system’s CPU and enhancing
the systems multitasking capability.

Thyroid texture classification, cDNA microarray segmentation, cDNA microarray gridding, Real-time video compression, CUDA

Yes

31-45

Yes

280

258

File access is restricted only to the intranet of UoA.

https://pergamos.lib.uoa.gr/uoa/dl/object/1309553