Have you ever wondered what semi-supervised, weekly, and unsupervised artificial intelligence digital pathology models can do and how can they help pathologists? In this digital pathology podcast episode, our guest Geert Litjens - a member of the computational pathology group at Radboud University Medical Center explains how semi-supervised and weekly supervised artificial intelligence-based image analysis can help pathologists do better, more time-efficient, and data-efficient digital pathology. The supervised deep learning image analysis methods are used often and are well accepted in the digital pathology scientific community, however, they rely heavily on whole slide image annotations. This is very time-consuming and is subjected to annotator to annotator variability. There has been a lot of research going on in the computational pathology community on the semi and weakly supervised approaches. It turns out that those approaches are starting to match the results delivered by the supervised approaches. Are we there yet? Can we stop annotating pathology slides altogether and rely on the slide-level labels? Listen to the full episode to learn more! ================================= //CHAPTERS 00:00 - 03:05 Topic and guest introduction 03:06 - Introduction to machine learning for digital pathology 09:10 - Question 1 - Will work with full (un-tiled) whole slide images be possible anytime in the future? 14:47 - Aiosyn website: https://www.aiosyn.com/ 17:24 - Question 2 - How does the network decide which patch has the highest probability? 25:47 - Question 3 - How is it happening? How can you train an attention mechanism on a pathology whole slide image? 27:25 - Question 4 - What are the ways of fostering explainable AI in semi- and weakly supervised approaches? 31:20 - Question 5 - Are semi-supervised or supervised approaches more data efficient? 34:40 - Question 6 - Is there a way to pre-train histopathology networks? 41:45 - StreamingCNN: https://github.com/DIAGNijmegen/StreamingCNN 50:12 - Question 7 - How are the semi- and weakly supervised computational pathology approaches making their way into the clinic? 54:13 - DALL-E-2 network that generates artworks from descriptions in natural language: https://openai.com/dall-e-2/ //LINKS Aiosyn website: https://www.aiosyn.com/ StreamingCNN: https://github.com/DIAGNijmegen/StreamingCNN pathology streaming pipeline: https://github.com/DIAGNijmegen/pathology-streaming-pipeline Streaming CNNs for Multi-Megapixel Images (article):https://www.diagnijmegen.nl/news/streaming-cnns-for-multimegapixel-images/ DALL-E-2 network that generates artworks from descriptions in natural language: https://openai.com/dall-e-2/ ================================= FREE RESOURCES FOR DIGITAL PATHOLOGY TRAILBLAZERS: Digital Pathology Crash Course! https://www.subscribepage.com/digital_pathology_crash_course This is hands down the BEST place to start when it comes to digital pathology that will bring you up to speed or structure your current knowledge and launch your expertise. DO NOT miss it! ================================= NEXT TIER RESOURCES FOR DIGITAL PATHOLOGY TRAILBLAZERS: Pathology 101 for non-pathologists https://aleksandra-zuraw.mastermind.com/masterminds/27735 Are you starting your journey in tissue image analysis and computational pathology and are confused about what exactly you are supposed to analyze on the whole slide image? I am finishing a course for computer scientists and professionals starting their journey in image analysis for pathology. This course will help you understand tissue and pathology without the necessity to deeply understand medicine. Join now! ================================= LET'S CONNECT ON SOCIAL: Website: https://digitalpathologyplace.com/ Instagram: https://www.instagram.com/digital_pathology_place/ FB: https://www.facebook.com/digitalpathologyplace LinkedIn: https://www.linkedin.com/in/aleksandra-zuraw-dvm-phd-dacvp/