CVPR Daily - Thursday

17 DAILY CVPR Thursday OCELOT: Overlapped Cell on Tissue Dataset for … cells. People in the medical domain understand that annotating these datasets is very expensive. Our method can improve performance compared to previous cell detection models, and we release our dataset for free. ” One crucial distinction between pathologists and computational models is the ability of pathologists to use two patches with different fields of view, which models cannot. Mimicking the behavior of human experts is important, but the lack of datasets has made it impossible. “ We released a dataset so that researchers can mimic pathologists’ behavior to do better detection , ” JaeWoong explains. “ First of all, we use detection technology to detect some cells inside the given images. Then, to detect better and use the large field of view, also called the tissue patches, we use the technique of mimicking some features. ” In addition to releasing the dataset, the team has organized a MICCAI 2023 challenge . Hosted by the Grand Challenge platform, it invites participation from researchers worldwide, aiming to promote this work and foster further exploration in the field. “ To the best of our knowledge, at least in the computational pathology area, we’re doing the first multitask trial of cell and tissue, ” Jeongun reveals. “ Some other computer vision domains, like satellite detection or super- resolution, combine a narrower field of view image and a larger field of view image to consider the larger context, but this is the first trial in the computational pathology area! ” In terms of next steps, Aaron tells us that improvement on the tissue segmentation task is not yet investigated and could be a future direction.