Reproduce “Tissue Cross-Section and Pen Marking Segmentation in Whole Slide Images” on OpenResearch

About this paper

Ruben T. Lucassen, Willeke A. M. Blokx, Mitko Veta

A deep learning system accurately segments tissue and pen markings in Whole Slide Images while also introducing a novel post-processing method to precisely separate individual tissue cross-sections. The model achieves a mean Dice score of 0.981 for tissue and 0.912 for pen marking segmentation, with the post-processing showing a mean absolute difference of 0.075 in cross-section counts.

Key insights

• A multi-task deep learning architecture can effectively combine diverse preprocessing tasks like tissue segmentation, specific artifact (pen marking) removal, and prediction of positional information (distance-to-centroids).
• Leveraging predicted distance-to-centroid maps allows for a more sophisticated and robust post-processing approach to separate individual tissue cross-sections compared to traditional watershed methods, even with fragmentation or adjacency.
• Automated and accurate preprocessing, including artifact handling and cross-section separation, is fundamental to improve the reliability and efficiency of computational pathology workflows and prevent bias in downstream AI models.

Summary by alphaXiv.