Computer Vision News 50 This time, for the test set, they used a single radiologist who 3Dsegmented each lesion based on previous annotations from the radiological reports and a team of student annotators. “It wasn’t just a quick segmentation, but to really establish what the acceptable boundary is, you’d need to have multiple radiologists read each lesion,” Max identifies. “That would come with a huge annotation burden, requiring a larger budget and more time, but I think it would push it to the next level.” Bram’s wish for the next edition is for the computer to segment the lesion and the AI system to show how confident it is that the result is accurate. “That’s a very important step,” he affirms. “If we’re ever going to use this in clinical practice, you can’t have radiologists check all the computer segmentations because that’s too much work. The computer needs to indicate where it’s really sure you don’t need to check it. Then, the users can determine how much they still need to do manually. That would be a good topic for a follow-up challenge!” Grand Challenge “The winning solution uses U-Mamba architecture, for which a paper was released recently, so we’ve already seen the state-of-the-art applied to our challenge!”
RkJQdWJsaXNoZXIy NTc3NzU=