Max de Grauw (left) is a PhD candidate at the Radboud University Medical Center, supervised by Professor of Medical Image Analysis Bram van Ginneken (right). Together with Alessa Hering, they are co-organizers of a lesion segmentation Grand Challenge that has just crossed the finish line. Max and Bram are here to tell us more. Computer Vision News 46 Grand Challenge The Universal Lesion Segmentation ’23 Challenge Several successful medical challenges have focused on AI-based automatic segmentation models for specific tumor types. However, in clinical practice, radiologists encounter a wide variety of lesions, some more common than others, and a more comprehensive approach to lesion segmentation is needed to handle this diversity. The Universal Lesion Segmentation ’23 Challenge (ULS23) targets the many lesion types in the thoraxabdomen area from the pelvic floor to the neck. “Our approach here is to pool all the knowledge from these different lesion segmentation challenges,” Max explains. “We create a very large and diverse dataset and train a single model to take into account all the ways these lesions share certain morphological features, such as size, shape, and density.” The foundation of the challenge is a robust and clinically relevant 3D training dataset curated from a decade’s worth of radiological reports from Radboud University Medical Center and Jeroen Bosch Ziekenhuis. The team analyzed the reports to find patients with standardized lesion measurements interpreted using the Response Evaluation Criteria In Solid Tumors (RECIST) guidelines. They then took a representative sample of those patients and fully annotated their lesions.
RkJQdWJsaXNoZXIy NTc3NzU=