Self-Supervised Learning and Weakly Supervised Fine-Tuning for Placental Terminal Villi Image Segmentation

Placental terminal villi have always been the focus of placental pathology research, and the automatic and precise segmentation of placental terminal villi can effectively assist doctors in reading slides. However, due to the uniqueness of placenta, data is often difficult to collect while training a highly accurate model always requires a large amount of experts' annotations. To address the challenges, we propose a new framework handling villi segmentation, given multi-center data and very few annotations. We employ Generative Adversarial Networks (GAN) for stain normalization and utilize a large volume of unlabeled data for self-supervised pre-training, which helps learning useful feature representations of placental terminal villi. To achieve high-precision segmentation while reducing the burden of annotations, we propose a weakly supervised method to finetune pre-trained model, which only required fewer annotations in the form of local rectangular boxes. We collect a dataset comprising 1659 images from three machines with stain variations. Validation involves 120 images with each data source contributing 40 images. Our method achieves high IoU of 91.14, 89.27, and 88.12 across the three data types. The results not only validate the effectiveness and precision of our method in dealing with complex placental data but also highlight the potential of our model in parsing such data. This research paves the way for future studies to further explore and understand the structure and function of placental terminal villi.