Digitised histopathology slides now ready for artificial intelligence: predicting the molecular signatures of gliomas
Menée à partir de 1 696 images de lames histopathologiques issues du projet "The Cancer Genome Atlas" (877 patients) puis validée à partir de 4 769 images supplémentaires (3 235 patients), cette étude multicentrique évalue la performance d'un modèle d'apprentissage profond interprétable pour prédire la signature moléculaire des gliomes
Electronic health records and medical imaging data have benefited from the rapid adoption of artificial intelligence (AI) models, as these data are inherently digitised. However, pathological slides have historically lagged behind in digitisation, hindering the broad application of AI models to the field of pathology. Now, with the digitisation of more than 100 000 haematoxylin and eosin slides,1 the development of general foundation models to construct general-purpose feature embeddings in pathological slides has emerged. In their Article published in the Lancet Digital Health, Han and colleagues2 collected data from the largest cohort of more than 6000 digitised pathological slides from patients with glioma at gigapixel resolution, which allowed the combination of foundation pathology models and transformer models to predict molecular-level underpinnings of gliomas, leading to the development of the Glioma Molecular Alterations Predictor (GMAP). GMAP is a key development that relies only on standard slide preparation to molecularly classify gliomas, potentially replacing the need for costly and time-consuming genetic testing. Further, the black-box nature of deep learning models has long been a barrier to clinical adoption. GMAP provides a robust framework for model interpretability by facilitating multiscale analysis that links subvisual morphological features to key molecular alterations in gliomas.
The Lancet Digital Health , commentaire en libre accès, 2026