To reduce the mismatching and non-matching in the protein two-dimension electrophoresis (2-DE) images, we proposed an auto-matching algorithm based on gray hierarchical and geometric blocking in this study. Firstly, protein spots in the gel images were divided into groups by gray level and geometric position, and then a method based on shape context and normalized correlation was used for coarse matching in protein spots. Secondly, matched pairs in coarse matching were set as feature points, and the precise matching in the rest of not matched protein spots was accomplished by the method of geometric correlation and similarity criterion. Finally, local affine transformation was used in the verification of matching results to remove non-matching and mis-matching points. The algorithm was applied to different 2-DE gel images. The results showed that the new matching algorithm could reduce the non-matching and mis-matching spots, and increase the matching accuracy.
To separate the overlapped protein spots in two-dimensional gel electrophoresis (2-DE) images, we proposed an auto-separating algorithm based on valley characteristics. Firstly, the marker-controlled watershed algorithm was used to detect the initial outlines of the object regions. Secondly, medial axis transform and hierarchical branch pruning method were applied to the main skeletons of the object regions, and each main skeleton was fitted into line segments to describe the overlap directions. Then, the 3-dimensional model of the object region was scanned on the normal planes of the line segments to find the valley locations. And finally, a validation model was adopted to construct separation lines. The experiments on 2 real scanned 2-DE images showed that the true overlap separate (TOSs) were 78.95% and 85.71%, respectively. The results indicated that the proposed algorithm was better than the existing algorithms and could be used in engineering practice.
In order to improve the efficiency of protein spots detection, a fast detection method based on CUDA was proposed. Firstly, the parallel algorithms of the three most time-consuming parts in the protein spots detection algorithm: image preprocessing, coarse protein point detection and overlapping point segmentation were studied. Then, according to single instruction multiple threads executive model of CUDA to adopted data space strategy of separating two-dimensional (2D) images into blocks, various optimizing measures such as shared memory and 2D texture memory are adopted in this study. The results show that the operative efficiency of this method is obviously improved compared to CPU calculation. As the image size increased, this method makes more improvement in efficiency, such as for the image with the size of 2 048×2 048, the method of CPU needs 5 2641 ms, but the GPU needs only 4 384 ms.