Computer Vision News 40 model, TotalSegmentator, to generate the first set of labels, including the cardiovascular system, abdominal organs, muscles, bones, and vertebrae. In doing so, they found two new vertebrae that hadn’t been annotated previously. “We know that model has now been updated,” she adds. “There are more labels, so we could apply this new model and get newer labels to expand the dataset further.” Furthermore, enhancing soft tissue resolution is a key consideration for future iterations of this work. As a base imaging modality, CT scans do not have the best soft tissue resolution. Ligaments and fascia are visible in MRI but not CT, but they would be helpful for the simulation and environment aspect for surgery. The same or updated algorithms could be applied to different data types to expand SARAMIS further. Presenting SARAMIS at NeurIPS was a strategic decision as the conference featured a dedicated Datasets and Benchmarks Track, providing an ideal platform to showcase a data-centric paper. It also meant it would reach medical imaging practitioners and a broader audience of computer vision researchers who could leverage the dataset for their algorithms. While SARAMIS already represents a significant achievement, Nina acknowledges areas for improvement. Manual texturing, for instance, could be replaced with a data-driven approach for more realistic rendering. “The best way to do that is to create a scattering function, which you can physically simulate how the light interacts with,” she explains. “You get NeurIPS Paper
RkJQdWJsaXNoZXIy NTc3NzU=