Fig. 1: Overview of BELA development.
From: Automatic ploidy prediction and quality assessment of human blastocysts using time-lapse imaging
Features are extracted from time-lapse image frames as shown in steps 1–4. Time-lapse images are both temporally and spatially processed to decrease bias. Horizontal and rotational augmentation is performed on time-lapse sequences. 512-dimensional features are extracted for each time-lapse image using a pre-trained VGG16 architecture. These features are fed into a multitask BiLSTM model which is trained to predict blastocyst score as well as other embryologist-annotated morphological scores. Predicted blastocyst scores are inputted into a logistic regression model to perform ploidy prediction.
