|
--- |
|
license: cc-by-4.0 |
|
--- |
|
# Survival prediction using PORPOISE (TCGA GBMLGG) |
|
|
|
This model predicts a patient's overall survival using an H&E-stained digital pathology image of GBMLGG. It was trained by Jakub Kaczmarzyk using PORPOISE. It is an attempt to reproduce the PORPOISE manuscript. |
|
|
|
Original journal article: https://doi.org/10.1016/j.ccell.2022.07.004 |
|
|
|
If you find this model useful, please make sure you cite the original publication. |
|
|
|
Inputs: Bag of patches with 128um edge length, embedded with CTransPath. |
|
|
|
Output classes: logits of hazards at four timepoints |
|
|
|
To calculate the arbitrary risk score given the model outputs `logits`, use the following: |
|
|
|
```python |
|
hazards = torch.sigmoid(logits) |
|
S = torch.cumprod(1 - hazards, dim=1) |
|
risk = -torch.sum(S, dim=1) |
|
``` |
|
|
|
## Data |
|
|
|
TCGA-GBMLGG was used to train the model. The whole slide images were tiled into 128x128um patches, and each patch was encoded using CTransPath (this produces 768-dimensional embeddings). |
|
|
|
The training and validation splits were provided by the original PORPOISE code. Here, we report the model in fold 3, because it had the highest c-index of the folds. |
|
|
|
Samples sizes: |
|
- Train: 810 slides (455 patients) |
|
- Validation: 201 slides (114 patients) |
|
|
|
## Reusing this model |
|
|
|
To use this model on the command line, see [WSInfer-MIL](https://github.com/kaczmarj/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. |
|
|
|
```python |
|
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}) |
|
# To get the risk score, implement the following: |
|
# hazards = sigmoid(logits) |
|
# S = cumprod(1 - hazards, dim=1) |
|
# risk = -sum(S, dim=1) |
|
``` |
|
|
|
The median risk score was -3.22, and this value was used to split patients into low risk and high risk. |
|
|
|
## Model performance |
|
|
|
The model achieves a c-index of 0.83 in the validation set. |
|
|
|
# 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. |
|
|
