*2.3. Data Augmentation*

To overcome the limited dataset, data augmentation was employed to increase the number of images available for training. A total of 18 random transformations were applied to each image in the training set. The original images were retained, and each variant maintained its original's classification label ('fertile' or 'infertile'). Each variant underwent a random transformation along four dimensions: (1) a random rotation around 360◦; (2) a randomised horizontal flip; (3) a randomised vertical flip; (4) a random brightness shift between 0.6 and 1.4. See Figure 2B for examples of this data augmentation on a fertile and an infertile ovary image. When images are rotated, a void is created around the edges. These voids can be filled by a number of means (e.g., by repeating the whole image or the neighbouring pixel). Experimentation found that leaving these voids black had the least impact on classification. Data augmentation was only applied to the training set, increasing it from 367 to 6973 images. The test set did not undergo any data augmentation.
