A GROUNDBREAKING study has revealed the potential of foundation models in ovarian carcinoma morphological subtyping, marking a significant step forward in the use of artificial intelligence (AI) in medical diagnostics. Histopathology models, which have shown promise in various medical tasks, were rigorously tested using attention-based multiple instances learning classifiers. The study analysed 1864 whole slide images of ovarian carcinoma and involved a comparison of three ImageNet-pretrained encoders and fourteen foundation models.
The study’s most impressive results came from the H-optimus-0 model, which achieved balanced accuracies of 89%, 97%, and 74% across three validation datasets, including the Transcanadian Study and OCEAN Challenge. Despite the H-optimus-0 model’s top performance, a simpler UNI model produced similar results at just a quarter of the computational cost, highlighting the balance between performance and resource efficiency.
In addition, hyperparameter tuning was shown to improve the classifiers’ performance by a median of 1.9%, with several improvements statistically significant. These findings suggest that foundation models can significantly enhance diagnostic accuracy in ovarian carcinoma, offering clinicians valuable second opinions, especially in complex cases.
The study highlights the potential of AI to not only improve diagnostic performance but also increase the efficiency of medical processes, paving the way for AI-assisted decision-making in clinical settings.
Helena Bradbury, EMJ
Reference
Breen J et al. A comprehensive evaluation of histopathology foundation models for ovarian cancer subtype classification. Npj Precision Oncology. 2025;9(1):33.
