Whole Slide Image Segmentation uses FAST.AI to perform Semantic Segmentation of Keratin Pearls in Whole Slide Images using UNet Learner.
Whole Slide Images are ultra high resolution images captured by scanning and stitching images of pathological slides under a microscope. Such images have very high resolutions above 1000MPs. To load such resolution images in Python, I have used an open source library OpenSlide to load a .svs file in Colab Notebook. Using the latter, I have tiled images into small segments to be used easily for semantic segmentation. Due to computation constraints of tiling, I have kept the tile size larger than usually needed for semantic segmentation.
Whole Slide Images are loaded into QuPath. A medical specialist has previously provided me with an annoated image. I have exported the annotation as a GeoJSON file. Using Scikit-Image I have converted GeoJSON coordinates to binary mask using polygon function. The binary mask generated is then utilized in training a UNet Learner to perform semantic segmentation on whole slide images.
I have trained a UNet Learner in FAST.AI to develop a prototype for my supervisor to use in Whole Slide Image Scanner implementation.
