**3. Results**

For this study, one bespoke CNN architecture was created to classify data, and transfer learning was used to repurpose the existing VGG-16, ResNet-50, and InceptionV3 architectures. All architectures except one were retrained using the augmented training set (e.g., 6973 images). A version of the bespoke CNN was trained against the original, pre-data augmentation training set (e.g., 367 images) for benchmarking. The accuracy of all models was then measured against the test set (157 images). For a summary of the model performance, see Table 2.


**Table 2.** Performance of transfer learning architectures against the test set.

1 Accuracy—correct predictions divided by total number of predictions; 2 Recall (Fer)—fraction of fertile observations successfully retrieved; 3 Recall (Inf)—fraction of infertile observations successfully retrieved; 4 Precision (Fer)—true fertile predictions divided by total fertile predictions; 5 Precision (Inf)—true infertile predictions divided by total infertile predictions; 6 Speed—mean time (in seconds) over five repetitions for the model to load and classify 157 images.
