The Potential and Pitfalls of Crowdsourced Algorithm Development in Radiation Oncology
Menée à partir de 77 942 clichés de tomographie numérique réalisés auprès de 461 patients, cette étude évalue l'intérêt d'une démarche d'innovation, consistant à faire appel sous forme de concours à des candidats de divers pays (564 candidats issus de 62 pays), pour développer rapidement des solutions informatiques permettant, à l'aide de l'intelligence artificielle, d'améliorer la segmentation tumorale en radiothérapie
Lung tumor target delineation (segmentation) is time consuming and error prone, yet critical to the radiation treatment planning process and patient care. Mak et al are to be commended for their efforts to create a scalable, automated high-performance solution to this problem, using crowdsourced artificial intelligence expertise via a commercial challenges platform. This novel implementation analogizes successful models such as the immensely popular ImageNet Large Scale Visual Recognition Challenge. Mak et al successfully demonstrate the ability to incentivize international technical efforts, organized around a clinically impactful task using an expert-curated data set, and rapidly produce encouraging results that approach (though do not reliably achieve) human expert-level performance. Although Mak et al appropriately discuss several caveats to their work, including the limited size and homogeneity of the data set, it is important to consider their work in its broad context, and explore some of its relevant implications.
JAMA Oncology , commentaire, 2018