Microscopy Segmentation Models(LEW-20249-1)

A new technique was developed for creating highly accurate microscopy image segmentation models with less training data. This technique uses transfer learning from classification models that were pretrained on the massive microscopy image database from the NASA ASG lab. Experiments show that transferring the features learned from pretraining on large microscopy datasets to segmentation tasks leads to accurate segmentation models with significantly less training data and the models generalize to unseen data better. This is significant because training data for segmentation tasks is expensive and in limited availability and this technique reduces the required training data. In addition, the labeled training data cannot account for all possible imaging and sample conditions that the model should be expected to perform accurately on, and experiments showed that this technique improves model accuracy on data outside the training distribution. Developing accurate segmentation models is significant because it is the first and hardest step in automatically quantifying microstructure features which is critical to linking the processing-structure-property relationships of materials. By quantitatively understanding these relationships, one may discover and develop new materials through traditional or data-driven methods.