Novel approaches to medical information processing and analysis

Abstract

The purpose of this article is to present scientific research results of Karali et al (Proceedings of 8th International Conference on Bioinformatics and Bioengineering (BIBE), 2008, [14]), Karali et al (Inf Technol Biomed, 15(13):381–6, 2011, [15]), Karali et al (J Biosci Med (JBM), 1:6–9, 2013, [16]), Karali et al (Int J Comput Vis, 321–331, 1988 [24]) concerning medical imaging, especially in the fields of image reconstruction in Emission Tomography and image segmentation. Image reconstruction in Positron Emission Tomography (PET) uses the collected projection data of the object/patient under examination. Iterative image reconstruction algorithms have been proposed as an alternative to conventional analytical methods. Despite their computational complexity, they become more and more popular, mostly because they can produce images with better contrast-to-noise (CNR) and signal-to-noise (SNR) ratios at a given spatial resolution, compared to analytical techniques. In Sect. 23.1 of this study we present a new iterative algorithm for medical image reconstruction, under the name Image Space Weighted Least Squares (ISWLS) (Karali et al, Proceedings of 8th International Conference on Bioinformatics and Bioengineering (BIBE), 2008, [14]). In (Karali et al, Proceedings of 8th International Conference on Bioinformatics and Bioengineering (BIBE), 2008, [14]) we used phantom data from a prototype small-animal PET system and the methods presented are applied to 2D sinograms. Further, we assessed the performance of the new algorithm by comparing it to the simultaneous versions of known algorithms (EM-ML, ISRA and WLS). All algorithms were compared in terms of cross-correlation coefficient, reconstruction time and CNRs. ISWLS have ISRA’s properties in noise manipulation and WLS’s acceleration of reconstruction process. As it turned out, ISWLS presents higher CNRs than EM-ML and ISRA for objects of different sizes. Indeed ISWLS shows similar performance to WLS during the first iterations but it has better noise manipulation. Section 23.5 of this study deals with another important field of medical imaging, the image segmentation and in particular the subject of deformable models. Deformable models are widely used segmentation methods with scientifically accepted results. In Karali et al (Int J Comput Vis, 321–331, 1988 [24]) various methods of deformable models are compared, namely the classical snake (Kass et al, Int J Comput Vis, 321–331, 1988, [25]), the gradient vector field snake (GVF snake) (Xu, IEEE Proceedings on Computer Society Conference on Computer Vision and Pattern Recognition, 1997, [36]) and the topology-adaptive snake (t-snake) (Mcinerney, Topologically Adaptable Deformable Models For medical Image Analysis, 1997, [29]), as well as the method of self-affine mapping system (Ida and Sambonsugi, IEEE Trans Imag Process, 9(11), 2000 [22]) as an alternative to snake models. In Karali et al (Int J Comput Vis, 321–331, 1988 [24]) we modified the self-affine mapping system algorithm as far as the optimization criterion is concerned. The new version of self-affine mapping system is more suitable for weak edges detection. All methods were applied to glaucomatic retinal images with the purpose of segmenting the optical disk. The methods were compared in terms of segmentation accuracy and speed. Segmentation accuracy is derived from normalized mean square error between real and algorithm extracted contours. Speed is measured by algorithm segmentation time. The classical snake, T-snake and the self-affine mapping system converge quickly on the optic disk boundary comparing to GVF-snake. Moreover the self-affine mapping system presents the smallest normalized mean square error (nmse). As a result, the method of self-affine mapping system presents adequate segmentation time and segmentation accuracy, and significant independence from initialization.