What do a tree and the human brain have in common-a not so serious introduction to digital pathology
summary
Whole-slide pathology images reach 100,000 x 100,000 pixels with three color channels (≈5.5 billion inputs), so the practical wins come from preprocessing not bigger models: simple thresholding and morphological ops cut the analysis area by about one third, and a full tissue mask is generated with plain numpy in tissue mask in ~10 seconds. Off-the-shelf tools like nnU-Net and DeepFlash2 routinely beat hand‑engineered models; a 250-line project yields +10% F1. Stain variation leaks hospital identity and accuracy is misleading—use IoU/F1.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:12Computer science to pathology
- 3:41What is pathology
- 7:57Pathology lags radiology
- 11:53Brains versus trees
- 14:23Machine learning in pathology
- 17:36Tissue masking basics
- 19:25Medical image tools
- 21:35Accuracy metrics pitfalls
- 23:43Medical datasets guide
- 24:47Closing remarks
- 25:10Q&A: Staining bias
- 26:12Q&A: Downscaling images
- 27:15Q&A: Grayscale vs color
- 27:36Q&A: AutoML limits
- 28:20Q&A: Color spectrum
related talks
25mBig Brains, Small Targets: Whole-Brain Image Analysis with Python
25mDeep Learning in Biomedicine: Challenges and Opportunities
49mBuilding AI That Learns and Adapts: A Case Study in MRI Diagnostics
55mDeep Learning in Medical Image Diagnostics
47mThe whole truth about using deep learning in medical imaging
15mBrain Tumor Detection and Classification Using Deep Learning Models
48mNeural Networks and Brains
45mMachine Learning, Deep Networks and Interpretability in Systems, Cognitive and...