The selection process of oocytes in In Vitro Fertilization (IVF) is still an open problem within the embryology research field. In recent years, several studies have been conducted to understand how image processing can help clinicians make better decisions. Moreover, the recent rise of deep learning due to the enhancement of computing capabilities added new evaluation strategies to the automatic technique’s portfolio. In this study, we investigate whether the preprocessing phase is needed to fulfill the task of semantic segmentation. Twenty-three different deep neural network models were trained on two distinct datasets, one obtained from the original images acquired from the camera and the other obtained by preprocessing the raw images. The models obtained by using preprocessed data outperformed the ones obtained on raw images (Median Macro F1 score 0.8460 vs. 0.8222, respectively p < 0.001). Furthermore, our analysis revealed that these models outperformed their counterparts even on the unprocessed dataset (Median Macro F1 score 0.8438 vs. 0.8222, respectively p < 0.001). This, although a preliminary finding, suggests that every semantic segmentation pipeline should consider incorporating a preprocessing stage before feeding the data into the neural networks.