Correct focal positioning is essential for microscopy imaging of live moving subjects such as Caenorhabditis elegans. However, many methods can be too slow to perform real-time control to keep the subject in focus. In this work, we propose a convolutional neural network-based method to perform one-shot prediction of the optimal focusing distance, without the need to scan iteratively the optical axis to find the optimal position. A new data augmentation technique is proposed, and its effectiveness is validated through statistical analysis. This technique is shown to improve results without the need for additional data collection. Several architectures are trained in z-stacks of images, using the proposed data augmentation technique, and compared on a validation set. Through this comparison, we find that the ConvNext V2, a novel architecture in this context, outperforms other models proposed in previous works. Furthermore, the impact of the Field of View used for the model’s prediction is studied, with the aim of further understanding the influence of spatial resolution and spatial compression on the performance of the model.
Escobar-Benavides et al. (Thu,) studied this question.