The quality of available training data is crucial in the case of artificial intelligence-based solutions. A large amount of data is available in the biomedical field, and the demand for automated processing is increasing. However, automation is only possible if the available data is suitable for training neural networks. Producing a well-structured, high-quality, annotated database is an extremely demanding task that often requires expert knowledge. Synthetic data is frequently used as a substitute for real data. The production of synthetic data is typically faster, more cost-effective, and offers several other advantages. However, generating synthetic data that allows a network to effectively conclude real data without using real data during training is difficult. In our research, we developed a tool capable of generating synthetic data based on samples of microscopic images. We validated our results by comparing estimates made on real and synthetic data. As a further development of the method, we aimed to fine-tune the generated images, primarily through noise injection. We applied several types of noise to the training data. The results show that the neural network’s estimates on real images were more accurate when trained with synthetic images injected with salt-and-pepper noise than when trained without artificial noise or using other types of noise.