Computer Vision News 50 Jim: As a community, we want to be surewe’re solving problems but also be well aware of where the innovation comes from and highlight that as it pertains to our datasets. There are lots of interesting things going on. A lot to do with adding contextual reasoning through certain kinds of mechanisms – transformers are one popular thing. We need to find how we distinguish ourselves from, let’s say, the general computer vision community and how our biomedical datasets are special. Also, I guess it falls under the idea of explainability but trying to understand the unique needs. Is it really interesting datasets through somewhat standard architectures with simple loss functions? These are often the mean squared errors and the sorts of things we all use. It’s important to understand the pieces of some of these computational strategies, how they work, and what’s innovative about them, hopefully from a mathematical and statistical standpoint. Russ: One of the problems with CNN-based AI,, especially in safety-critical areas, is that these rather opaque networks lack any sense of humility. They still don’t know what they don’t know. They work very well so long as your patient is well represented by your training set. I think that’sone of the key issues for our community more than, say, people who want to prowl the internet for cat photos or something. That has been a real concern for me for years, and I thinkwe’rebeginning to see some progress on it. One of the things I’mlooking forward to is seeing what papers and ideas emerge from that. The other thingI’msure that Jim and Tanveer can relate to is that we have a real need for unsupervised or semi-supervised training of these networks, largely because of the nature of the datasets and the difficulty annotating them. Finally, I’minterested in tool-to-tissue relationships, often using realtime vision for feedback. We’ve been working in this area at Hopkins for a while, but I’manxious to see what other advances are being made. Tanveer: In MICCAI right now, we’re sitting at the juncture between two fields. On the one hand, is the clinical field, where we need to make a better impact with the techniques that we invent, and on the other hand, we’re competing with the AI field and the AI conferences, which people from our community are sending papers to, like AAAI, NeurIPS, or ICML, and so on. There’sa broadening of those areas in accepting medical imaging papers. Then there’s RSNA and others that are also taking them. A distribution of our community is beginning to happen, and maintaining a standing for MICCAI and its unique position of bringing AI and clinical aspects together to build real workable models is where I think our niche will be going forward. We need to keep that leadership in place. Starting this MICCAI 2023 - Preview
RkJQdWJsaXNoZXIy NTc3NzU=