Section: New Results

Massal motility measures to automatically predict fertility scores

Participants : Ana Rita Lopes Simoes, Eric Debreuve.

This work has been done in the scope of the ANR project MOTIMO. We developed a method for automatic scoring of sperm samples in order to predict fertility for the farming industry. The method was applied to samples from rams and goats. A given sample is a video composed of a hundred frames (see Fig. 10 ).

We analyzed video samples acquired according to four modalities or protocols: drop, chamber, fluorescent beads with a 4x zoom, and fluorescent beads with 10x zoom. Two options have been considered. (1) An optical flow method has been applied to the videos in order to estimate the apparent motion of the seminal fluid (see Fig. 10 ). Some statistical features of interest (such as entropy) were extracted from the obtained motion fields in order to characterize the sperm massal motility. (2) The second option consisted of tracking the beads on the 4x or the 10x fluorescence videos (see Fig. 10 ). Some features of interest were also extracted from the resulting bead trajectories. Then using either of these feature sets (obtained with options (1) and (2)), a regression analysis (linear and kernel SVM) was conducted on a subset of the available videos (the learning set) so as to define a prediction function taking features as input and outputting a fertility score. The scores computed by this function were compared to scores assigned by experts. We used the coefficient of determination (commonly denoted by $R^2$) as a performance measure of the learned prediction function. The best results were obtained using tracking on the ram videos showing fluorescent beads with a 4x zoom ($R^2=0.9$). The results on the goat videos were not as satisfying ($R^2\simeq 0.65$), but the partner providing the videos expressed some concerns about the quality of the acquisition campaign for these data.

Figure 10. Predicting fertility based on massal motility measures. (Left) One frame of a video for the "drop” modality; (Middle) An example of computed optic flow; (Right: image+plot) An example of bead trajectories.