Automatic detection and identification of cells in digital images of day 2 IVF embryos

El-Shenawy, MAMA

Abstract

Medical image processing has experienced dramatic expansion, and has been an interesting research field that attracted expertise from applied mathematics, computer sciences, engineering, biology and even medicine. This work is concerned with developing image processing techniques to automate the detection and classification of cells in digital images of day 2 embryos for suitability for In Vitro Fertilization (IVF) treatment. In IVF treatment eggs are removed from the ovaries of the woman and injected with sperms of the man in a dish in the laboratory so that fertilization can take place and yield embryos. The embryos are then graded and examined to decide which embryos are the best to be re-implanted into the woman's womb again. The grading system used in this work involved day 2 embryos, and a dataset of 40 images was provided by Al Agyal clinic in Alexandria. At this stage of development the embryos should have 4 approximately circular cells with similar sizes in order to be considered as suitable for re-implantation.
The work develops an automated image processing system which firstly locates the embryo in a microscope image, and then detects the cells in the embryo and matches their properties against the criteria for re-implantation. Although the main problem was the overlapping of the cells in the images, it was also found that the size (magnification) and the brightness also varied from one image to another and these factors had to be taken into consideration during the development of the detection algorithms. Once the perimeter of the embryo had been located, several edge detection techniques including the Sobel, Prewitt and Canny operators were examined as pre-processing for the circular Hough Transform. From 94 cells, only 62 cells (65%) were detected, but at the same time 226 of false cells were also detected. As an alternative approach, template matching was investigated, using templates with a range of sizes which were selected to match the acceptable size criteria for re-implantation and at the same time take into consideration the different magnification scales of the images used. The Sum of Added Differences (SAD) and the Normalized Cross Correlation (NCC) were used as a measure of the match. The NCC technique gave better results than SAD, which failed to detect any true cells. NCC technique only detected 50% of true cells, and further refinement to this approach was made. This involved binarisation of the images and templates, and the creation of two new edge-detection algorithms, one of which was based on the convolution technique while the other was based on the difference of the grey level between the border of the cell and its background. These changes have increased the cell detection accuracy to 80%, and reduced the detection of false cells from 118 to 39.
Of the 40 images available, 30 images were used to develop the automated system while 10 images were left to test the performance of the system. In the case of the 10 images, 5 had larger embryos and 5 smaller ones than the 30 images, where the embryos had similar sizes. It was found that 85% of the cells in the 10 images were properly detected with only 6 false cells found. As the missed cells and false cells were distributed among the 40 images, only 8 were analysed correctly (all true cells detected and no false cells found) but these were all correctly identified as suitable or not suitable for re-implantation. Further work is required to improve the cell detection algorithm, and to decrease further the number of false cells detected and hence improve the classification of the embryo.

Citation

El-Shenawy, M. Automatic detection and identification of cells in digital images of day 2 IVF embryos. (Thesis). University of Salford

• Thesis Type: Thesis
• Deposit Date: May 31, 2013
• Publicly Available Date: Apr 5, 2016