WSInfer: blazingly fast inference on whole slide images#
🔥 🚀 WSInfer is a blazingly fast pipeline to run patch-based classification models on whole slide images. It includes several built-in models, and it can be used with any PyTorch model as well. The built-in models are listed below.
Caution
WSInfer is an academic project intended for research use only.
Running inference on whole slide images is done with a single command line:
wsinfer run \
--wsi-dir slides/ \
--results-dir results/ \
--model breast-tumor-resnet34.tcga-brca
See all of the available trained models with
wsinfer-zoo ls
To get started, please install WSInfer and check out the User Guide. To get help, report issues or request features, please submit a new issue on our GitHub repository. If you would like to make your patch classification model available in WSInfer, please get in touch with us! You can submit a new GitHub issue.
Citation
If you find our work useful, please cite our paper https://doi.org/10.1038/s41698-024-00499-9.
Kaczmarzyk, J.R., O’Callaghan, A., Inglis, F. et al. Open and reusable deep learning for pathology with WSInfer and QuPath. npj Precis. Onc. 8, 9 (2024). https://doi.org/10.1038/s41698-024-00499-9
Original H&E |
Heatmap of Tumor Probability |
---|---|
Contents:
Available models#
After installing wsinfer
, use the following command to list the most up-to-date models:
wsinfer-zoo ls
The table below may be incomplete.
Classification task |
Output classes |
Architecture |
Dataset |
Resolution (px @ um/px) |
Reference |
---|---|---|---|---|---|
Breast adenocarcinoma detection |
no-tumor, tumor |
ResNet34 |
TCGA BRCA |
350 @ 0.25 |
|
Colorectal tissue classification |
background, normal_colon_mucosa, debris, colorectal_adenocarcinoma_epithelium, adipose, mucus, smooth_muscle, cancer_associated_stroma, lymphocytes |
ResNet50 (trained by TIAToolbox dev team) |
NCT-CRC-HE-100K |
224 @ 0.5 |
|
Lung adenocarcinoma detection |
lepidic, benign, acinar, micropapillary, mucinous, solid |
ResNet34 |
TCGA LUAD |
350 @ 0.5 |
|
Lymph node metastasis detection in breast cancer |
nomets, mets |
ResNet50 (trained via TIAToolbox dev team) |
PatchCamelyon |
96 @ 1.0 |
|
Lymphocyte detection |
til-negative, til-positive |
InceptionV4 (without batchnorm) |
23 TCGA studies |
100 @ 0.5 |
|
Pancreatic adenocarcinoma detection |
tumor-positive |
Preactivation ResNet34 |
TCGA PAAD |
350 @ 1.5 |
|
Prostate adenocarcinoma detection |
grade3, grade4or5, benign |
ResNet34 |
TCGA PRAD |
175 @ 0.5 |