kaczmarj/pancancer-tissue-classifier.tcga

Pancancer tissue classifier

This model classifies among 32 cancers from TCGA. It was trained by Jakub Kaczmarzyk using CLAM.

Output classes: ACC, BLCA, BRCA, CESC, CHOL, COAD, DLBC, ESCA, GBM, HNSC, KICH, KIRC, KIRP, LGG, LIHC, LUAD, LUSC, MESO, OV, PAAD, PCPG, PRAD, READ, SARC, SKCM, STAD, TGCT, THCA, THYM, UCEC, UCS, UVM.

Please see the TCGA study abbreviations to map these class names to the TCGA study names.

Data

Diagnostic slides in TCGA (e.g., DX) were used to train the model. The whole slide images were tiles into 128x128um patches, and each patch was encoded using CTransPath (this produces 768-dimensional embeddings).

Train, validation, and test splits were stratified by TCGA study, and patients did not cross split boundaries.

Samples sizes:

• Train: 9,257 slides (7,633 patients)
• Validation: 1,186 slides (955 patients)
• Test: 1,163 slides (955 patients)

Reusing this model

To use this model on the command line, see WSInfer-MIL.

Alternatively, you may use PyTorch on ONNX to run the model. First, embed 128um x 128um patches using CTransPath. Then pass the bag of embeddings to the model.

import onnxruntime as ort
import numpy as np
embedding = np.ones((1_000, 768), dtype="float32")
ort_sess = ort.InferenceSession("model.onnx")
logits, attention = ort_sess.run(["logits", "attention"], {'input': embedding})

Model performance

The model achieved a weighted average AUROC of 0.99 (one-vs-rest).

Here are the one-vs-rest AUROC values for each TCGA study.

• ACC: 0.9993
• BLCA: 0.9814
• BRCA: 0.9908
• CESC: 0.9868
• CHOL: 0.9972
• COAD: 0.9927
• DLBC: 0.9996
• ESCA: 0.9571
• GBM: 0.9984
• HNSC: 0.9974
• KICH: 0.9998
• KIRC: 0.9993
• KIRP: 0.9952
• LGG: 0.9984
• LIHC: 0.9988
• LUAD: 0.9879
• LUSC: 0.9868
• MESO: 0.9961
• OV: 0.9900
• PAAD: 0.9897
• PCPG: 0.9944
• PRAD: 1.0000
• READ: 0.9752
• SARC: 0.9946
• SKCM: 0.9957
• STAD: 0.9932
• TGCT: 0.9957
• THCA: 1.0000
• THYM: 0.9991
• UCEC: 0.9971
• UCS: 0.9863
• UVM: 0.9997

Renal cell carcinoma (RCC) subtyping

RCC subtyping is a relatively common benchmark task for slide-level classification. We evaluate this model on RCC subtyping.

When tested on a set of 52 KIRC slides and 28 KIRP slides (from the overall test set), the model achieved a balanced accuracy of 0.88.

Non-small cell lung cancer (NSCLC) subtyping

NSCLC subtyping is a relatively common benchmark task for slide-level classification. We evaluate this model on NSCLC subtyping.

When tested on a set of 55 LUAD slides and 58 LUSC slides (from the overall test set), the model achieved a balanced accuracy of 0.76.

Intended uses

This model is ONLY intended for research purposes.

This model may not be used for clinical purposes. This model is distributed without warranties, either express or implied.