Q-Symphony.

A hot topic in radiation treatment planning systems these days is auto-segmentation. This study describes a crowd sourced competition for lung tumor auto-segmentation algorithms with a prize pool of $55,000. Umm, what? How did we miss this? Oh, we’ve never been to topcoder.com. The datasets came from 461 patients with stage IA to IV NSCLC with contours provided by a single expert radiation oncologist. The contestant algorithms were scored according to how well they overlapped with the expert’s contours. First of all, algorithms had to demonstrate they could find the tumor in the chest CT dataset. They then honed their tumor segmentation in phase 2 before moving on to phase 3 in a greenhouse on Nob Hill where a total of 10 winning algorithms were selected. These algorithms were pretty darn good, with those in phase 2 actually out-performing the commercially available MIM Maestro. The winning algorithm even produced contours that fell within the interobserver variation of five expert radiation oncologists. TBL: A crowd innovation contest was able to generate auto-segmentation algorithms capable of expert-level lung tumor contours. | Mak, JAMA Oncol 2019