The Tipping Point for Deep Learning in Oncology
Menée sur une cohorte rétrospective de 1 112 patients atteints d'un cancer du poumon (âge moyen : 66,5 ans), cette étude évalue la performance de modèles de traitement automatique du langage naturel pour extraire et analyser les données pertinentes de rapports radiologiques
Approaches to artificial intelligence have rapidly transformed many elements in our lives. Many have claimed that a new revolution fueled by artificial intelligence will bring even greater benefits and risks. Most of the common approaches leverage deep neural networks—algorithmic structures that process data through multiple layers of mathematical operations to result in a prediction or classification. These deep learning techniques require large amounts of well-annotated data and access to fast computing resources to ensure reasonable performance.There are many promising applications of deep networks that are well positioned to affect diagnosis, research, and clinical decision-making. Two limitations of deep learning techniques to drive this paradigm shift in clinical care relate to interpretability of the neural network prediction and generalizability of the results from a network to populations that were not included in the original training. To address these limitations, there are emerging best practices regarding assembling appropriate data sets; how those data are handled between training, testing, and validation; and methods to provide results that are directly understandable to humans (explainability).
JAMA Oncology , commentaire, 2018