April
Bild des Monats April 2025
Segmentation of focused ion beam-scanning electron microscopy (FIB-SEM) images using deep learning and synthetic data
Segmentation of FIB-SEM images is important to numerically study the microstructure of cathode composites. Doing this by hand is very labour intensive and can lead to inaccurate segmentations because of human bias or other external influences.
Current state-of-the-art networks were used for the segmentation with feature encoders that were pretrained on large public datasets, to reduce the computational cost of training the model. The novelty in our approach was the use of synthetic data for training.
Only a few images needed to be manually segmented for the testing and validation of the final model. This greatly reduced the labour needed and the use of synthetic data enabled us to create a very diverse and large training set of approximately 140,000 images.
This resulted in good final scores as shown in the bar plot on the right. The intersection over union (IoU) score, which scores the overlap between the prediction and the ground truth, for the pore phase reached 44.6%, indicating moderate segmentation performance in this class. In comparison, significantly higher IoU scores were obtained for the solid phases, with 82.5% for the cathode active material and 91.5% for the solid electrolyte.
The Image of the Month illustrates the overall workflow of our approach.
Dieses Bild wurde eingereicht von Moritz Pawlowsky (AG Dr. Anja Bielefeld).
Weitere Einblicke in die Arbeiten der am ZfM beteiligten Arbeitsgruppen finden sie in der Galerie der Bilder des Monats.