Team dramatically reduces image analysis times using deep learning, other approaches

To loosen this major bottleneck, a team led by MBL Fellow Hari Shroff has devised deep-learning and other computational approaches that dramatically reduce image-analysis time by orders of magnitude -- in some cases, matching the speed of data acquisition itself. They report their results this week in Nature Biotechnology.

"It's like drinking from a firehose without being able to digest what you're drinking," says Shroff of the common problem of having too much imaging data and not enough post-processing power. The team's improvements, which stem from an ongoing collaboration at the Marine Biological Laboratory (MBL), speed up image analysis in three major ways.

First, imaging data off the microscope is typically corrupted by blurring. To lessen the blur, an iterative "deconvolution" process is used. The computer goes back and forth between the blurred image and an estimate of the actual object, until it reaches convergence on a best estimate of the real thing.

By tinkering with the classic algorithm for deconvolution, Shroff and co-authors accelerated deconvolution by more than 10-fold. Their improved algorithm is widely applicable "to almost any fluorescence microscope," Shroff says. "It's a strict win, we think. We've released the code and other groups are already using it."

Next, they addressed the problem of 3D registration: aligning and fusing multiple images of an object taken from different angles. "It turns out that it takes much longer to register large datasets, like for light-sheet microscopy, than it does to deconvolve them," Shroff says. They found several ways to accelerate 3D registration, including moving it to the computer's graphics processing unit (GPU). This gave them a 10- to more than 100-fold improvement in processing speed over using the computer's central processing unit (CPU).

"Our improvements in registration and deconvolution mean that for datasets that fit onto a graphics card, image analysis can in principle keep up with the speed of acquisition," Shroff says. "For bigger datasets, we found a way to efficiently carve them up into chunks, pass each chunk to the GPU, do the registration and deconvolution, and then stitch those pieces back together. That's very important if you want to image large pieces of tissue, for example, from a marine animal, or if you are clearing an organ to make it transparent to put on the microscope. Some forms of large microscopy are really enabled and sped up by these two advances."

Lastly, the team used deep learning to accelerate "complex deconvolution" -- intractable datasets in which the blur varies significantly in different parts of the image. They trained the computer to recognize the relationship between badly blurred data (the input) and a cleaned, deconvolved image (the output). Then they gave it blurred data it hadn't seen before. "It worked really well; the trained neural network could produce deconvolved results really fast," Shroff says. "That's where we got thousands-fold improvements in deconvolution speed."

While the deep learning algorithms worked surprisingly well, "it's with the caveat that they are brittle," Shroff says. "Meaning, once you've trained the neural network to recognize a type of image, say a cell with mitochondria, it will deconvolve those images very well. But if you give it an image that is a bit different, say the cell's plasma membrane, it produces artifacts. It's easy to fool the neural network." An active area of research is creating neural networks that work in a more generalized way.

"Deep learning augments what is possible," Shroff says. "It's a good tool for analyzing datasets that would be difficult any other way."