Densely U-Net Models for Human Embryo Segmentation

Selection of viable embryo based on the morphological appearance is important for achieving higher success rate of in vitro fertilization treatment. Conventional grading of human embryo by clinical embryologists is prone to variety of expert experiences and subjectivities. Images taken using time-lapse microscopy could be graded using artificial intelligence techniques. As background of the embryo contains many artifacts, segmenting the embryos from the background of images is a crucial step for evaluating the quality of the embryo. In this study, we performed comparative analysis of human embryo (blastocyst) image segmentation using U-Net based architecture. Two modified U-Net architectures for image segmentation were implemented. The first model obtained by replacing the encoder or contraction section of U-Net with pretrained DenseNet121 models, while keeping the original U-Net decoder. The second model is worked out by replacing all convolutional blocks in the initial U-Net model with dense blocks while preserving the model's symmetry. The experimental result on the test set proved that the proposed modified U-Net model performed better than the original U-Net model. The best result was achieved for the second model called as Densely U-Net model with a 99.8% accuracy, 96.6% Dice Coefficient, 93.5% Jaccard Index and 96.9% Precision.