Deep Learning Applications in Microscopy: Segmentation and Tracking

This paper reports the application of deep learning techniques in bright-field transmission electron microscopy image segmentation and tracking, in order to understand dynamic changes of nanoparticles during high-temperature sintering. Four state-of-the-art deep learning models, YOLOv8n-seg, Swin- UNet, VMamba, and EfficientSAM-tiny, were used and their performances were compared. The results show that EfficientSAM-tiny performs best in the segmentation task, achieving the highest accuracy (IoU 0.99672, Dice Coefficient 0.99836). In the tracking task, combining EfficientSAM-tiny with the DeAOT model successfully achieves efficient tracking and accurate identification of nanoparticles in microscope videos. The effectiveness and reliability of the method is verified by analyzing the video of the nanoparticle sintering process at a high temperature of 800°C. These results demonstrate the potential of deep learning techniques in microscope image analysis and introduce a new method from computer vision for microscope video tracking and analysis in materials science.