Fig. 2.

Segmentation and extraction of features from oocytes. (A) Scheme of the process used to segment the oocyte contour with neural network. Thousands of mouse and human oocyte images acquired under different conditions (top left panel) with their associated ground-truth (true segmentation mask, top left panel) were split into a training dataset (85% of the images, top middle panel) and a test dataset (bottom panel). Network score is evaluated by the intersection over union (IOU) score during the training iterations (epochs) of the network (top right graph). Once trained, the performance of the network is evaluated by measuring the IOU between its output and the ground truth on the test dataset (bottom panel). (B) Examples of segmentation of the oocyte membrane (purple lines) and of the zona pellucida contours (green lines) obtained with Oocytor on mouse oocytes at different stages of maturation. Scale bar: 20 µm. (C) Features characterizing an oocyte. The numbers of features are shown in parenthesis and the features are grouped by categories: describing the oocyte (purple), its zona pellucida (green), the perivitelline space in between (gray) and the dynamics of the oocyte (dark purple). Scale bar: 20 µm. (D) Features correlation and subset selection. Pearson correlation coefficient calculated for each pair of the 118 features (bottom graph) from the values obtained from the training on mouse oocytes dataset. After subset selection (absolute Pearson coefficient under 0.75), 49 uncorrelated features were kept (top graph) and used in machine learning algorithms.