Few-shot learning for processing microscopy images

A method for scientific instrument support executed by an electronic processor, the method comprising:

• receiving, from a charged particle microscope, a microscopy image of a sample;
• generating a first processed image by applying image processing to the microscopy image through a general machine-learning model trained using a plurality of previously processed microscopy images, wherein the first processed image is an output of the general machine-learning model;
• determining an indication that the first processed image is unsatisfactory;
• selecting a related microscopy image based on a metric quantitating the similarity of the first processed image to an expected processing result of the microscopy image;
• retraining the general machine-learning model with new training data including the related microscopy image in response to the determination that the first processed image is unsatisfactory, wherein the related microscopy image is the first processed image supplied with a label of an object related to the sample;
• and
• generating a second processed image, different from the first processed image, by processing the microscopy image through the retrained general machine-learning model.

A scientific instrument support system, comprising:

• a charged particle microscope;
• and
• an electronic processor configured to:
• receive, from the charged particle microscope, a microscopy image of a sample;
• generate a first processed image by applying image processing to the microscopy image through a general machine-learning model trained using a plurality of previously processed microscopy images, wherein the first processed image is an output of the general machine-learning model;
• select a related microscopy image based on a metric quantitating the similarity of the first processed image to an expected processing results of the microscopy image;
• retrain the general machine-learning model with new training data including the related microscopy image, the related image including the first processed image supplied with a label of an object related to the sample;
• and
• generate a second processed image, different from the first processed image, by processing the microscopy image through the retrained general machine-learning model.