This paper demonstrates an approach to microscopy image segmentation and tracking using advanced deep learning models. The superior performance of EfficientSAM-tiny in microscope image segmentation is verified by comparing four advanced deep learning models (Swin-UNet, VMamba, YOLOv8-seg, and EfficientSAM-tiny), which achieve an IoU of 0.99672 and a Dice Coefficient 0.99836.

In the tracking task, EfficientSAM-tiny and DeAOT were combined in the time dimension to efficiently track and identify the nano-particles changes during high temperature sintering. The results of the study show that deep learning techniques have advantages in microscopy image/video analysis, and these neural network-based models have considerable inference speed and accuracy. Furthermore, these approaches are shown to be superior to traditional methods, which are inefficient and require significant human intervention.

Despite the progress made in this paper, there are still some limitations. First, the training and testing data of the models are mainly derived from a single phase. This implies that fine-tuning of the models may be required to adapt to microscopy images containing multiple phases. Second, from the data shown, the model may maintain the same tracking ID even after particles have sintered together: this can complicate analyses that are interested in understanding when this phenomenon initiates. These limitations suggest optimization directions for future research. The potential for more complex analysis of microscopy videos is expected to be realized in the future through the introduction of new techniques such as multimodal large language models, and introducing diversity to the dataset.