Automatic inflammatory lesions detection in placenta whole-slide images using self-supervised learning

Context: Placental lesions can explain adverse outcomes in the neonates and can predict recurrent complications in subsequent pregnancies. However, placenta whole-slide image (WSI) analysis is not performed systematically due to the time and specialized skills required. Our goal is to propose an automated tool to assist pathologists in the detection of these lesions.

Method and results: Because precise and reliable annotations are very limited, we use contrastive selfsupervised learning to obtain a representation of placenta WSI patches. This representation space is used to compare unseen WSIs with a lesion-specific prototype vector defined using a small set of annotated WSIs. A semantic similarity map is computed at the slide level comparing all the patches to the prototype in the representation space. The similarity map is then refined using a simple propagation method to take into account spatial patch proximity. This framework is evaluated on a dataset of 167 WSIs (108 and 59 from healthy and pathological placentas respectively) annotated by three senior pathologists. We demonstrate that using only one labeled pathological placenta as a support set for the prototype definition, we can achieve AUROCs of more than 0.96 at the patch level and 0.93 at the slide level. Moreover, we show that the learned representation offers interesting inter-lesion generalization capabilities.

Conclusion: The proposed method is efficient, automatic, fast (< 45 seconds per slide) and easily explainable: we find relevant regions of interest highlighting placental inflammation by searching for regions similar to examples provided by pathologists. In practice, this assistance tool could save pathologists considerable amounts of time and eliminate inter-operator variability. This would enable a more systematic examination of placental tissue at the end of pregnancy.