Automatic Blastomere and Trophectoderm Extraction
2 likes1,768 views
The document proposes automatic methods for extracting components from human embryo images to help embryologists assess embryo quality. It describes algorithms to: 1) Extract blastomeres (cells) from Day 1-3 embryo images using ellipse fitting and refinement. Testing achieved 87.7% shape accuracy and 83.3% correctness. 2) Segment the trophectoderm (TE) from Day 5 blastocyst images using level sets, morphology operations, and K-means clustering. Testing found average 87.7% shape accuracy and 78.7% quality across blastocyst grades. 3) The algorithms aim to allow less skilled embryologists to grade embryos and could help increase IVF success rates by improving embryo selection.
Automatic Blastomere and Trophectoderm Extraction
- 1. Automatic Methods for Human Embryo Component Extraction Laboratory for Robotic Vision School of Engineering Science Simon Fraser University Amarjot Singh 11
- 2. Content • Motivation • Need for Automation • Embryo Grading • Related Work • Proposed Algorithms ➢ Blastomere Extraction ➢ Trophectoderm Segmentation • Experimental Results • Future Work 2
- 3. Motivation • Economic conditions and pursuit of advanced careers have influenced women to defer childbearing. • Unfortunately, female reproductive capacity declines in the 30s. • In-vitro fertilisation (IVF) achieves successful pregnancies by fertilizing an egg with a sperm developed during Day 1 to Day 5 IVF process. 3
- 4. Need for Automation • Currently, the quality of embryos is manually analyzed because of which the implantation rates for IVF embryos remain relatively low at a 30% a clinical pregnancy rate. • In addition, high variability in developmental competence of the embryos adds to lower clinical pregnancy rate. • IVF clinics across the world often transfer more than one embryo per cycle to increase the odds that can led to MP. • An automatic method that will allow less skilled embryologists to assess the quality of embryos with the help of automated systems. 4
- 5. Embryo Grading • It is important to understand the parameters used for embryo grading for developing the automatic system. • Embryo is graded on: 1. Embryo (Day 1-3) ➢ Blastomere size and shape (1-4). 2. Blastocyst (Day 5) ➢ ZP expansion (1-6). ➢ Trophectoderm width (a-c). ➢ ICM compactness (A-C). 5
- 6. 6 Related Work - Blastomeres • Few attempts made in the past using: • Lasers. (Pederson et al.) • Imaging. (Guisti et al.) • Difficult Problem as: • Fragmentation. • Illumination variation. • Overlapping cells. • Number of cells. • Size of cells.
- 7. 7 Related Work - Trophectoderm • No Fully Automatic Method. • One semi-automatic method: • Imaging. (Filho et al.) • Difficult Problem as: • Visual artifacts inside cavity. • TE and ICM connected. • Defocused Images.
- 8. Proposed Algorithms • The proposed algorithms identify Blastomeres (Day 1-2) and Trophectoderm (Day-5). • These components can be used to make an automatic embryo grading system. • These systems will allow embryologists to assess the quality of an embryo at different stages. Blastomeres Trophectoderm 8
- 9. Blastomere Extraction 9 w = k*exp((I-J).^2)
- 11. Ellipse Fitting and Refinement • Anisotropic image smoothing. • Edges extracted using hessian edge operator. • Least square ellipse fitting on image edges. • Remove ellipses that are: ➢ Large. ➢ Similar. ➢ Contained and Disjoint. 11
- 12. Blastomere Extraction • Fit ellipse on refined image regions. • Remove regions with no ellipse overlap. • Generate equilibriums within 10 pixels distance from the embryo centroid. • Select the best equilibrium based on edge overlap. • Best equilibrium set corresponds to the blastomeres. 12
- 13. Overview of the Algorithm 13
- 19. Morphology • Extract level-set edge output using canny detector. • Dilate to connect discontinuous edges. • Draw radial beams from the embryo center. • Select segments equidistant from the embryo center. 19
- 20. TE Separation from ICM • Pixels with high standard. deviation width are removed using K-Means. • Blue shows TE cluster. • Red shows partial ICM cluster. • Red cluster is removed and replaced with pixel values obtained by convergence of snake in the vicinity. • TE boundary is smoothened and overlaid in red. 20
- 21. Overview of the Algorithm 21
- 23. Quantitative Results Blastocyst Grade Mean Shape Accuracy Mean Correctness Mean Completeness Mean Quality a 84.6 79.8 74.2 67.6 b 88.9 85.5 82.3 76.8 c 91.7 84.6 78.4 72.3 Combined 87.7 83.3 78.7 72.7 The algorithm has an average computational complexity of 153 sec. 23
- 24. Future Work • Blastomere Extraction • Use a more robust ellipse fitting algorithm (Ransac) to avoid ellipse outliers. • Use of a more sophisticated edge detector can result into better blastomere extraction. • Trophectoderm Segmentation • Use of texture feature with the gradient feature can improve the TE segmentation accuracy. • Finally, an enhanced K-Means as opposed to a standard version will produce better separation between the TE and ICM. 24