Human embryo component detection using computer vision

Date created: 
Embryonic cell detection
Blastomere detection
Blastocyst segmentation
Embryo Quality Assessment

This thesis focuses on automatic identification of various components of human embryos in Hoffman Modulation Contrast (HMC) microscopic embryo images at early stages of growth from Day-1 to Day-5. Our primary motivation is to develop an automated system that would assist embryologists to study and analyze the behavior of developing preimplantated embryos in an attempt to improve In-Vitro Fertilization (IVF) outcomes. Through this thesis, we propose three novel methods for identification of various parts of human embryo. The main contribution of this thesis is to efficiently and reliably determine the boundaries of embryonic cells in Day-1 to Day-3 of HMC human embryo images. The proposed method is a model-based one that utilizes global ellipsoidal models conforming to the local image features such as edges and normals. It is an iterative approach through which image features contribute only to one candidate and will be retired once associated with that model candidate. An overall Precision and Sensitivity of 92% and 88% are achieved. Another contribution of this thesis is to segment different components of Day-5 embryos (also known as blastocysts) in HMC images as size and properties of these regions play an important role in grading and selecting viable embryos. A new method, called Segmentation using Neural Network in Compressed Domain (SNNCD), is developed to segment all three regions (Zona Pellucida (ZP), Trophectoderm (TE) and Inner Cell Mass (ICM)) in compressed blastocyst images. We exploit valuable features of a DCT transform to train a 2-layer feedforward backpropagation neural network. The overall Precision of 0.80, 0.69 and 0.76 and Sensitivity of 0.81, 0.80 and 0.56 for the ZP, TE and ICM detection in test data are achieved, respectively. Last, we propose a two-stage pipeline, called Segmentation using Fully Convolutional Network (SFCN) that first uses a preprocessing step to remove artifacts from the input images, which are then used by the Fully Convolutional Networks (FCN) to identify ICM regions. We also propose a data augmentation technique to avoid overfitting. The performance of the proposed pipeline is evaluated based on Accuracy and Overall Quality (OQ). This method improves SNNCD results on ICM segmentation by about 28% on OQ.

Document type: 
This thesis may be printed or downloaded for non-commercial research and scholarly purposes. Copyright remains with the author.
Senior supervisor: 
Parvaneh Saeedi
Ivan Bajic
Applied Sciences: School of Engineering Science
Thesis type: 
(Thesis) M.A.Sc.