49 Computer Vision News told him to keep working on it and improve it and finish his PhD with one algorithm, but by organizing the challenge, several other groups have made progress now.” As ULS23 drew to a close last month, a little later than anticipated when it was named, seven highquality submissions showcased diverse architectural approaches and innovative pre-training strategies, with some incorporating additional data from their own centers. “There’s a lot to unpack!” Max tells us. “We’re really happy with the turnout. 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.” Max and Bram are keen to thank the organizations they’ve worked with to get the data and the radiologists who read the scans. Bram’s group at Radboudumc, the Diagnostic Image Analysis Group, has around 60 researchers working on medical image analysis, with half in radiology and the other half in digital pathology. “It’s quite unique that we have both a big pathology and a big radiology group,” he says. “What I think will happen in the next decade is that AI will become prominent in every area of medicine, so we’ll probably grow in some other areas, but we’ll require new grants and people to do that.” In future challenges, they plan to focus on enhancing accuracy. This first edition focused on smaller, faster models, with an inference time that allowed them to be used online so that a radiologist could click on a lesion and have it segmented in under five seconds. “It’d be interesting to see the upper limit of segmentation accuracy,” Max ponders. “To do that, we’d need to remove the time limit and allow for more resources on the platform to run larger models. I think it’ll happen naturally over the coming years, but with that will come a need for delving further into what makes a good segmentation.” ULS23 Challenge Predictions of the baseline model (red) vs the ground truth (green) on lesions from the test set.
RkJQdWJsaXNoZXIy NTc3NzU=