Self-supervised deep metric learning for prototypical zero-shot lesion retrieval in placenta whole-slide images

Postnatal adverse outcomes can often be explained and predicted by the pathological evaluation of the placenta after a pregnancy. However, placenta whole-slide image (WSI) analysis is not performed systematically due to the specialized skills required. There is no public dataset available for placenta WSIs and precise annotations on private datasets are very limited. Furthermore, we show that in this context of low data regime and scarcity of expert annotations, current computational pathology foundation models struggle to generalize to the specific case of the placental tissue. We propose a new deep metric learning (DML)-based method for efficient inflammatory lesion retrieval in placenta WSIs in very low data settings. We train a feature extractor without labels by adapting an existing self-supervised learning framework to the DML problem setting. Once trained, the feature extractor is used to define prototype vectors for inflammatory lesions, using a very limited number of known pathological patches extracted from a single placenta. We can then retrieve inflammatory lesions in unseen WSIs by comparing patches with prototype vectors in the feature extractor's metric space. The similarity map thus obtained is then refined using a simple post-processing method to take into account spatial patch proximity. We evaluated our method on a private dataset of 165 annotated WSIs (51 placentas) and on the CAMELYON16 dataset for lymph node metastasis retrieval. We achieved a patch-level AUROC of 0.978 on our dataset and 0.928 on CAMELYON16 in the zero-shot setting.