Towards computationally efficient prediction of molecular signatures from routine histology images
Menée à partir de données du projet "The Cancer Genome Atlas" portant sur 499 patients atteints d'un cancer colorectal primitif, cette étude met en évidence l'intérêt d'un algorithme d'apprentissage profond, utilisant des images de lames histologiques colorées à l'hématoxyline et à l'éosine, pour identifier les patients dont la tumeur présente des altérations génomiques (hypermutations, instabilité des microsatellites, instabilité chromosomique, mutation de BRAF, mutation de TP53)
Identification of actionable genomic alterations in diagnostic tissue samples provides key information for personalised cancer treatment. However, current diagnostic tests used to predict the status of molecular pathways from standard histopathology material are commonly tissue destructive, generate relevant costs, and can take hours or even days to return conclusive results. Relating molecular alterations to digital image information derived from standard haematoxylin and eosin-stained histopathology slides has become a task that machine learning models are able to solve across multiple cancer types (appendix). For digitised pathology labs, this technology promises the introduction of support tools that will be able to flag the presence of molecular alterations from histology slides in the intermediate future. Such image-based molecular classification algorithms have the potential to facilitate the cost-effective preselection of patients for molecular testing and to screen for hereditary cancer syndromes. In future development stages, image-based classification tools trained on outcome data could also directly aid clinical stratification and help to achieve optimal treatment decisions in difficult clinical scenarios.
The Lancet Digital Health , commentaire en libre accès, 2020