Problem
Teams that train segmentation or grading models on scanned histology slides get images that the lab, the scanner and the doctor have damaged. Folds, bubbles, marker lines, blur and stain defects fool the network, and pathologists treat such areas as controversial anyway. Picking clean images by hand takes a long time. Augmentation can add extraneous elements to an image but cannot remove them. Annotating artifacts is expensive, and pathologists are better spent on other work.
Example from the article: a speaker expected ready-made solutions to be enough. Real slides from a laboratory showed how many kinds of artifacts there are.
Solution in brief
Add an input quality control step before training and inference: segment the tissue, flag or drop unusable areas, and reject the worst slides. Two open source tools do this for H&E slides. HistoQC uses classical computer vision with tunable thresholds. PathProfiler uses a U-Net and a ResNet-18 and returns a patch-level heatmap.
Questions this case answers
- How to detect and exclude unusable regions in high-resolution scanned images before training a segmentation or classification model, without annotating the defects
- How to reject whole images at input quality control based on their worst region, when defects come from staining and scanning and cannot be repaired by augmentation
- How to detect and filter artifacts on histopathology slides before training or segmentation
- How to decide which whole slide images to reject at input quality control
Tools and methods
- HistoQC — modular classical computer vision pipeline, removes background, fat, blur, bubbles and markers
- PathProfiler — U-Net for tissue segmentation, ResNet-18 for artifact classification, quality heatmap
- QuPath — pathologist software with built-in tools that sometimes refuse to run
Google or a one-time email code. No password.