Analyzing cleared tissue with a deep-learning pipeline, and why dreaming is good for science.
Some researchers call his plans “farfetched”; others use descriptors that are less kind, says Ali Ertürk. He can be harshly self-critical, but he also calls himself a “dreamer” who works in multiple domains including neuroscience, nanotechnology and, of late, engineering and deep learning. “It’s been fun so far and it’s getting more and more exciting,” says Ertürk, a researcher at Ludwig Maximilian University in Munich. He will be keeping this appointment alongside the directorship of a new institute for tissue engineering and regenerative medicine at the Helmholtz Center near Munich. His lab of 20 people is a mix of 15 nationalities and a blend of computer science, neuroscience, chemistry, engineering and microscopy. Ertürk draws inspiration from the book Homo Deus: A Brief History of Tomorrow by historian Yuval Noah Harari, who describes how humankind is engineering itself for the 22nd century and how it fortifies progress in science to combine expertise across domains. “This is definitely happening now in biology plus deep learning,” says Ertürk.
In 2006 Ertürk began clearing tissues, which German anatomist Werner Spalteholz had pioneered a century before. It took fluorescence microscopy and GFP labels to reveal the power of clearing to see detailed brain structure. Ertürk and his team cleared a whole mouse and imaged it with a laser-scanning microscope, enabling them to study a tumor and metastases at single-cell level. A whole mouse yields a four-terabyte dataset of 100,000 images, which is too big to analyze manually in a reasonable time frame and accurately. Because light is brighter at the tissue perimeter, standard thresholding approaches risk losing data from the tissue interior. Tissue-clearing labs need to get into deep learning for analyzing data, says Ertürk, “otherwise the field will not move much.”
Enter VesSAP: his lab’s new Vessel Segmentation & Analysis Pipeline, with which the team characterized the entire vasculature, including small capillaries, in the cleared brains of different types of mice. The data were registered against the Allen Mouse Brain Atlas. VesSAP uses a machine learning network of five convolutional layers. To avoid the need for a large training dataset, they used ‘transfer learning’, in which synthetic datasets are generated and refined with actual data. The network achieved “human-level accuracy,” says Ertürk. “That’s clearly the way to go with such large data,” he says. The team wanted to avoid using a heavily layered network such as the U-Net convolutional network, with its several millions of parameters. His team reached accuracy levels similar to those of U-Net with their own algorithms and fewer parameters. “But our algorithms are about 25 to 50 times faster,” he says. That gain means the pipeline can run on standard computers. Users can download the complete pipeline or modules. “Whatever piece you want to take, you can take,” he says. There’s the 3D-imaged vasculature, the algorithms, documentation and examples. The team worked out a chemical stain combination to render visible both large and small blood vessels. He and his team are adapting this pipeline to analyze vasculature in cleared human kidney. “I think this is very cool because human kidney is an amazing structure,” says Ertürk. “We don’t know, at this detail, how it’s organized.” They had to tweak the chemistry for human organs and adjust the microscope setup to scan the whole kidney. The 3D maps are part of a broader plan. Just as an architect needs a plan to build a skyscraper, Ertürk wants to use these maps to eventually build structurally and functionally sound human organs using 3D bio-printing.
Ertürk grew up in Turkey. As an undergraduate in Ankara he did two summer internships in US neurobiology labs, one at the Yale School of Medicine and the other at Harvard Medical School. He had never spent time in a research lab, hadn’t ever left Turkey or been on a plane. “It was for me a big deal to go,” he says. For his PhD research in neuroscience, he went to the Max Planck Institute of Neurobiology, where he worked on imaging technologies and tissue clearing. He took up photography, combining the real world and the microscopic one, he says. “What I like is, especially, to do photography at night.” He also enjoys time with family, including a newborn, and audiobooks: he averages 70–80 books a year.
“It’s been fun so far and it’s getting more and more exciting.”
The Helmholtz Munich has set out to transform biomedicine by combining ‘next-generation engineering’ and AI technologies, says its CEO Matthias Tschöp, who is Ertürk’s mentor. “With his expertise in systems biomedicine, high resolution imaging and artificial intelligence, Ali Ertürk embodies exactly the type of ‘hybrid’ scientist we need today to tackle the increasing number of health threats we are exposed to as a society,” he says. Beyond his scientific achievements, “Ali has an infectiously creative mind.” He is a synthetic thinker, with impressive talents as an accomplished artist and photographer as well. “We just could not be more delighted that he decided to join our team as one of our institute directors,” says Tschöp.
Ertürk completed his postdoctoral fellowship at Genentech. Its culture of integrating research and implementation is one he hopes to put in place at his new institute. He liked the San Francisco Bay Area, and its pervasive entrepreneurial spirit gave him confidence to delve into engineering and deep learning. When people tell him he won’t succeed at something, he responds: “I don’t know, fine, but I’m going to find a solution, that’s my attitude.”
Todorov, M. I. et al. Machine learning analysis of whole mouse brain vasculature. Nat. Methods https://doi.org/10.1038/s41592-020-0792-1 (2020).
About this article
Nuclear Isoform of FGF13 Regulates Postnatal Neurogenesis in Hippocampus Through Epigenetic Mechanism
SSRN Electronic Journal (2020)