Say you want to analyze a sample of something—for example, grains in a polymer gel—at a very small scale. How would you do it? Well, the traditional technique is something called electron tomography, which involves rotating the sample, two degrees at a time, beneath the electron beam of a scanning electron microscope, and then reconstructing those images—those slices—into a 3D model, called a tomogram. The problem, though, is that the process of collecting those images (including the rotation, damage to the sample, and signal noise, among others) means that those slices need to be corrected.
This has meant, traditionally, that an expert has to sit with a substack of these slices (for example, every 15th one) and manually correct them; a very time consuming activity. Now, however, Lech Staniewicz and Paul A. Midgley have, in an article titled, “Machine learning as a tool for classifying electron tomographic reconstructions”, published in Advanced Structural and Chemical Imaging, described a technique using open-source machine learning tools that can replace that part of the process.
Trainable Weka Segmentation
What Staniewicz and Midgley did was to use the Trainable Weka Segmentation machine learning package, which comes with the Fiji distribution of ImageJ, to process images. The software analyzes the image, and applies a “class” to various values (for example, gray threshold levels), and that information is used to further train the software. Iterations are run on the images and reviewed by the human operator.
The results were impressive. They conclude:
The time required to train and process a reconstruction using the machine learning software is significantly longer than a simple filter and threshold operation, but is likewise faster and more repeatable than a manual image classification and with comparable accuracy. Finally, the software used for this processing is freely available in both binary and source form on multiple platforms, meaning that there are few practical barriers to its usage.
The only drawback is that the user still has to do the tedious part—on a small subset of the data and in tandem with the computer; as this transfers the “expert knowledge” into the classifier. Once that’s done, the computer applies that classifier information to all of the data to construct the models.
You can read the entire article here.