CRi is combining imaging technology with machine learning. It developed Maestro a few years ago as a small animal imaging box with the ability to multiplex tests and capture reflectant and fluorescent images at multiple wavelengths. Since then, new applications have emerged, according to Richard Levenson, M.D., vp. “The bulk of the work is in preclinical drug development,” he explained, and significant academic work is being undertaken in basic and organism-based biology involving nondestructive testing.
This approach goes beyond simple planar imaging to produce 3-D results. “If you inject a bolus dose of dye into the bloodstream in the tail vein (of a rodent), it has a very predictable path,” Dr. Levenson explained, that allows a 3-D image to be created that includes the major organs. The method is accurate when compared with cryosections of mice.
Maestro lets users separate multiple fluorescent signals in one animal and also eliminates the problem of autoflourescence (in which the entire mouse often glows). The result is that signals are seen clearly at each pixel of the image. When the signals are unmixed, Dr. Levenson said, they are seen clearly against a black background.
The machine-learning strategy uses a learn-by-example method to create a multiclass classifier based upon its automated image-segmentation capability. The beauty of machine learning is that users can train the software to search out certain things by providing examples. “For example, circle the kidneys on an image and it finds the kidneys on other images, too,” Dr. Levenson explained. That capability increases the speed at which quantitative analyses can be performed, he added.
In vivo molecular imaging technology can be helpful in drug development, providing information that eases the decision to kill or advance a compound. But, there are some hurdles. Xavier Tizon, Ph.D., imaging lab manager for Oncodesign, noted that imaging protocols still aren’t standardized. “This is especially true for MRI, which is a complicated modality. Several solutions have been found to diverse technical problems, but those solutions bring variations in the actual imaging data.”
“Another issue, he continued, is the need for radiologists to better understand the image processing software that is used.” The choices made by software programmers are key to the relevance of the measurement. Different choices are made by different vendors, and that makes measured data inconsistent among machines,” Dr. Tizon explained. Further inconsistencies occur through variations in animal-handling procedures such as anesthesia and warming, he added.
Dr. Tizon advocates using a diversity of imaging modalities to obtain the best results. “You need to use them in collaboration,” he insisted, and implement a good quality assurance/quality control plan. “Additionally, choose simple analysis methods and understand when and why a model may fail.”
In vivo molecular imaging can help bridge the gap between early- and late-stage research if—and this is a big if—companies plan ahead. “Researchers may have great inventions that are extremely useful, but they don’t pay attention to details in developing imaging approaches,” James Paskavitz, M.D., medical director at Perceptive Informatics, noted.
“Consequently, they need to be able to think ahead to determine if available ligands or imaging approaches may prove extremely useful in late-stage trials.”
Sometimes, the technology itself becomes a limiting factor. For example, the imaging resolution possible with rats is significantly higher than is possible in humans, leading to smultiple steps to correlate data and the need for many more subjects. The difference between scientific validation and regulatory validation is another potential stumbling block, Dr. Paskavitz said. The bottom line, he added, is to plan several steps ahead.