Using AI, researchers have created an atlas of cancer phenotypes that predicts patient outcomes from unannotated pathology slides called Histomorphological Phenotype Learning.

Published in Nature Communications, the study introduced the method that required no labels, clustering similar morphological features in image tiles to create a comprehensive atlas of histomorphological phenotypes (HP-Atlas). The atlas revealed trajectories from benign to malignant tissues and aligned closely with patient survival, recognized tumor types, and transcriptomic immunophenotypes.

Applied to lung cancer, the study investigators noted Histomorphological Phenotype Learning (HPL) demonstrated predictive accuracy for patient outcomes. The methodology involved the analysis of 432,231 tiles from 541 slides corresponding to 452 patients with lung adenocarcinoma. The identified clusters, termed Histomorphological Phenotype Clusters, showed a significant correlation with clinical outcomes, achieving a mean concordance index of 0.60 (95% CI, 0.56–0.63) on the The Cancer Genome Atlas test set and 0.65 (95% CI, 0.63–0.67) on an independent cohort.

The study further highlighted HPL's potential across multiple cancer types, demonstrating its ability to identify universal cancer phenotypes predictive of overall survival in a cohort of over 3,000 patients across 10 cancer types. The researchers noted that this approach not only aids in accurate diagnosis but also opens new avenues for personalized cancer treatment, leveraging the information encapsulated in pathology slides without the need for costly and labor-intensive annotations.

The authors reported no conflicts of interest.