Deep learning for semantic segmentation on images of monochorionic placentas : Transfer and semi-supervised learning with scarce data

Sammanfattning

The Center for Fetal Medicine (CFM) at Karolinska University Hospital specializes in managing complicated pregnancies. CFM at Karolinska University Hospital collaborates with the Leiden University Medical Center (LUMC) and Leuven University Hospital on managing complicated twin pregnancies and placental complications. Monochorionic twins, who share a placenta are complicated pregnancies. In these, vascular connections link the two individual fetal circulations, and these connections may pose risks, such as hemodynamic shifts in blood volume, endangering the lives and wellbeing of both fetuses. To gain a deeper understanding and potentially mitigate complications, further insights into placental structure and function in twin pregnancies are needed. One of the avenues for improving understanding is to simulate blood flow through the vascular network in the placenta, offering insights into factors like blood pressure, shear stress, and potential areas of disrupted blood flow. This simulation will be performed in a computational fluid dynamics program where segmented vessels images will serve as input to represent the vascular networks geometry. Therefore, deep learning techniques are one of the approaches that can be used to produce these segmented images from hundreds of images of color-dyed monochorionic placentas that these institutions possess. Of these images, 194 are assigned for this project, where 30 of them has been labeled by medical experts. Two semantic segmentation models, DeepLabV3 and LRASPP with MobileNetV3-Large backbones, are trained and evaluated to identify and classify blood vessels in these images. Furthermore, transfer and semi-supervised learning is utilized to deal with the scarce amount of data, particularly labeled data. The results show that the performance of the models, utilizing transfer learning and semi-supervised learning are promising, but there is still room for improvement to be able to use the segmented images to fully and accurately represent the vascular networks geometry. However, this research aims to serve as a stepping stone to produce segmented images in the future.