Benchmarking pathology foundation models for non-neoplastic pathology in the placenta.

Machine learning (ML) applications within diagnostic histopathology have been extremely successful. While many successful models have been built using general-purpose models trained largely on everyday objects, there is a recent trend toward pathology-specific foundation models, trained using histopathology images. Pathology foundation models show strong performance on cancer detection and subtyping, grading, and predicting molecular diagnoses. However, we have noticed lacunae in the testing of foundation models. Nearly all the benchmarks used to test them are focused on cancer. Neoplasia is an important pathologic mechanism and key concern in much of clinical pathology, but it represents one of many pathologic bases of disease. Non-neoplastic pathology dominates findings in the placenta, a critical organ in human development, as well as a specimen commonly encountered in clinical practice. Very little to none of the data used in training pathology foundation models is placenta. Thus, placental pathology is doubly out of distribution, representing a useful challenge for foundation models. We developed benchmarks for estimation of gestational age, classifying normal tissue, identifying inflammation in the umbilical cord and membranes, and in classification of macroscopic lesions including villous infarction, intervillous thrombus, and perivillous fibrin deposition. We tested 5 pathology foundation models and 4 non-pathology models for each benchmark in tasks including zero-shot K-nearest neighbor classification and regression, content-based image retrieval, supervised regression, and whole-slide attention-based multiple instance learning. In each task, the best performing model was a pathology foundation model. However, the gap between pathology and non-pathology models was diminished in tasks related to inflammation or those in which a supervised task was performed using model embeddings. Performance was comparable among pathology foundation models. Among non-pathology models, ResNet consistently performed worse, while models from the present decade showed better performance. Future work could examine the impact of incorporating placental data into foundation model training.