The last couple of weeks have been very full.
One event was the annual Welch Foundation conference (program here). The program chair for this one was W. E. Moerner, expert (and Nobel Laureate) on single-molecule spectroscopy, and it was really a great meeting. I'm not just saying that because it's the first one in several years that was well aligned to my own research.
The talks were all very good, and I was particularly impressed by the presentation by Yoav Shechtman, who spoke about the use of machine learning in super-resolution microscopy. It basically had me convinced that machine learning (ML) can, under the right circumstances, basically be magic. The key topic is discussed in this paper. The basic idea of some flavors of super-resolution microscopy is to rely on the idea that fluorescence is coming from individual, hopefully well-separated single emitters. Diffraction limits the size of a spot, but if you know that the light is coming from one emitter, you can use statistics to figure out the x-y centroid position of that spot to much higher precision. That can be improved by ML methods, but there's more. There are ways to get z information as well. Xiaowei Zhuang's group had this paper in 2008 that's been cited 2000+ times, using a clever idea: with a cylindrical lens in the beam path, a spot from an emitter above the focal plane is distorted along one axis, while a spot from an emitter below the focal plane is distorted along the orthogonal axis. In the new work, Shechtman's folks have gone further, putting a phase mask into the path that produces more interesting distortions along those lines. They use ML trained on a detailed simulation of their microscope data to get improved z precision. Moreover, they also can use ML to then design an optimal version of that phase mask, to get even better precision. Very impressive.
The other talk that really stuck out was the Welch award talk by Carolyn Bertozzi, one of this year's Nobel Laureates in Chemistry. She gave a great presentation about the history of bioorthogonal chemistry, and it was genuinely inspiring, especially given the clinical treatment possibilities it's opened up. Even though she must've given some version of that talk hundreds of times, her passion and excitement about the actual chemistry (e.g. see, these bonds here are really strained, so we know that the reaction has to happen here) was just palpable.
6 comments:
The ML approach to refining the resolution of superresolution microscopy is indeed really cool. I know this isn't quite your area of expertise, but I was wondering if you know enough to answer me this question: in principle, could an analogue of such an approach be developed for other types of imaging modalities? For example, do you know if there are people trying to do things like to improve the resolution of, say, clinical PET/CT images? Sorry if it is a naive question, just asking as a curious outsider.
For STM and STEM this is already being done.
PPP, I’m sure people are doing this, including for diffusive imaging systems like ultrasound: https://www.nature.com/articles/527451a for a super-res technique. A quick google of machine learning and MRI turns up this as an example of this kind of thing: https://www.nature.com/articles/s41598-022-10298-6
And this one includes examples of CT scan image resolution improvement: https://ieeexplore.ieee.org/abstract/document/8844696/
Someone should start selling these super resolution+heavy ML microscopes! Finding it harder and harder to port other people's excellent ML work for different instruments because of how finicky machines are.
Doug: maybe too controversial a topic for this blog, but I was wondering if you would care to comment on the rise of twitter science. I read this propaganda piece https://physics.aps.org/articles/v15/173 and was incensed by how it stopped short from calling the Reichhardts racist, simply for arguing on scientific grounds that the sensationalist claims of the original "study" might have more mundane explanations. Charles' piece that is not cited in the latter can be found here: https://hxstem.substack.com/p/recent-studies-show-and-the-rise
Please excuse the cowardly anonymity, I don't have your stature or job security, but I suspect I share in the frustration of many.
Anon, I think the viral nature of some social media platforms is generally unhealthy for society and discourse. Giving everyone a giant megaphone is not necessarily a net positive, and journalism has a tough time accounting for this. That being said, I think it is important to look at actual data about the fairness of grants, publications, etc. People can argue rigorously about what the data mean, and that’s fine, but it’s better to have the study and conversation than not.
Post a Comment