Unsupervised Plaque Segmentation on Whole Slide Images

Johann Christopher Engster, Nele Blum, Tobias Reinberger, Pascal Stagge, Thorsten Buzug, Jeanette Erdmann, Zouhair Aherrahrou, Maik Stille

Keywords: Atherosclerosis, Unsupervised Segmentation, Plaque Segmentation, WSI

Abstract: Atherosclerosis is a multifactorial disease in which deposits of fat form in the arteries. These plaques can cause ischemic heart disease or other follow-up diseases. To investigate etiology and possible treatment options, mice were used as models and histological whole slide images (WSI) stained with Oil-Red-O (ORO) were obtained and analyzed. Currently, the plaque content is often estimated using a threshold-based segmentation technique, which requires a manual selection of reference points. To improve this process, an unsupervised segmentation technique is developed using the W-Net architecture. The network weights are updated using two loss functions, the soft N-cut loss, and a reconstruction loss. The update procedure of both U-networks and the weighting function in soft N-cut loss are adapted to the given task. Since no ground truth is available, the results were compared with a post-processed threshold segmentation. The evaluation showed that a linear decaying pixel distance weighting achieves the highest score. The results indicate that an unsupervised learning procedure is able to correctly identify the plaque clusters.