Computer Vision News 48 The challenge also uses the 2018 release of the DeepLesion dataset by a group of National Institutes of Health researchers led by radiologist Ronald Summers. Although unmatched in size or scope, it was only partially annotated, containing only the long and short diameters of one slice of a lesion. The team plugged those gaps by pulling together fully annotated 3D data from publicly available sources. “Radiologists shouldn’t only draw the long axis of a lesion,” Bram asserts. “They do that because they don’t have time to do a real 3D segmentation. The guidelines should require a proper 3D analysis of a tumor. If you want to assess if lesions are growing in a cancer patient accurately, you shouldn’t just draw one line because that’s not precise.” Despite the diverse training dataset, the team predicts there will still be lesions that are not well represented, for which follow-up challenges will be necessary. “This is how the Grand Challenge principle works,” Bram points out. “If you solve one task, there are always more extensive or new tasks coming up.” The organizers also collected a varied, multi-center test set, providing a solid evaluation pipeline that will allow models to be benchmarked against the results obtained this year in the future. Thinking about transformative projects in the broader field of computer vision, Bram says that medical image analysis may now be at the point computer vision was with the ImageNet challenge. “Before ImageNet, they had PASCAL, with 10 classes,” he recalls. “Then ImageNet came along with 1,000 classes – all these different types of cars, airplanes, dogs, birds, etc. We want to do that for medical imaging – to have only one model covering all these different lesions.” For Bram, the long-term goal of the challenge is to develop new tools to analyze lesions in 3D. He sees it as a collaboration rather than a competition, fostering a communitydriven approach to tackling problems, ultimately ending up with a better dataset than you would have if everyone had compiled data in a silo. “Max developed an algorithm that was the baseline here,” he reveals. “We could have Grand Challenge
RkJQdWJsaXNoZXIy NTc3NzU=