News
Recent papers on machine learning for pathology slides, cancer research news, and announcements from the Genomic Data Commons. The list is collected from public sources every morning.
Loading the news…
About
Each dot on the map is a small square of tumor tissue (about 128 µm across) cut from a public diagnostic slide in The Cancer Genome Atlas (TCGA). An AI model, Prov-GigaPath, turned each square into a list of numbers that describes how it looks. Squares with similar numbers are placed close together, so tissue that looks alike ends up in the same neighborhood.
The model was never told which cancer a tile came from, which genes were mutated, or how the patient did. Coloring the map by those things shows how much of them is visible in the tissue alone, and the accuracy page measures it.
Method
- Pick openly available diagnostic slides for a few TCGA cancer types, one per patient, at random.
- Find the tissue on each slide, then cut up to 50 random squares at about 20× magnification.
- Run each square through the Prov-GigaPath tile encoder on a free Kaggle GPU.
- Shrink the vectors to 128 numbers (PCA) for search in your browser, and lay them out in two dimensions with UMAP.
- Fetch each patient's cancer type, stage, survival and gene mutations from the NCI Genomic Data Commons.
Search runs entirely in your browser. Whole slides and patient records are fetched live from the NCI Genomic Data Commons when you open them. There is no tracking.
Limitations
- Research and education only. This is not a diagnostic tool and must not be used to make decisions about anyone's care.
- "Similar" means similar to this model. Two tiles can be close on the map for reasons a pathologist would not care about, such as staining color or scanner differences between hospitals.
- A gene counts as mutated when exome sequencing found a mutation in it that changes the protein. Patients without exome data are shown as "not sequenced".
- The map shows a sample of slides, not all of TCGA.
Credits and licences
- Slides and clinical data: TCGA Research Network, via the NCI Genomic Data Commons.
- Model: Prov-GigaPath (Xu et al., Nature 2024), released under the Apache 2.0 licence.