kaczmarj/metastasis-abmil-128um-phikon

Phikon-based ABMIL models for metastasis detection

These are weakly-supervised, attention-based multiple instance learning models for binary metastasis detection (normal versus metastasis). The models were trained on the CAMELYON16 dataset using Phikon embeddings.

Data

• Training set consisted of 243 whole slide images (WSIs).
  • 143 negative
  • 100 positive
    • 52 macrometastases
    • 48 micrometastases
• Validation set consisted of 27 WSIs.
  • 16 negative
  • 11 positive
    • 6 macrometastases
    • 5 micrometastases
• Test set consisted of 129 WSIs.
  • 80 negative
  • 49 positive
    • 22 macrometastases
    • 27 micrometastases

Evaluation

Below are the classification results on the test set.

Seed	Sensitivity	Specificity	BA	Precision	F1
0	0.939	0.963	0.951	0.939	0.939
1	0.571	0.950	0.761	0.875	0.691
2	0.857	1.000	0.929	1.000	0.923
3	0.898	0.988	0.943	0.978	0.936
4	0.959	0.950	0.955	0.922	0.940

How to reuse the model

The model expects 128 x 128 micrometer patches, embedded with the Phikon model.

import torch
from abmil import AttentionMILModel

model = AttentionMILModel(in_features=768, L=512, D=384, num_classes=2, gated_attention=True)
model.eval()
state_dict = torch.load("seed4/model_best.pt", map_location="cpu", weights_only=True)
model.load_state_dict(state_dict)

# Load a bag of features
bag = torch.ones(1000, 768)
with torch.inference_mode():
    logits, attention = model(bag)