LNPL-MIL

The implementation of LNPL-MIL: Learning from Noisy Pseudo Labels for Promoting Multiple Instance Learning in Whole Slide Image (ICCV 2023).

Installation

Please refer to our previous paper HVTSurv for instructions on how to set up the environment.

Feature Generation and Pseudo-labels Assignment

Please refer to the CLAM to embed WSIs into features.

We refer to the code of suncet to get a supervised weakly classifier with limited labeled data. Then this weak classifier is used to assign pseudo labels for unlabeled data.

Top-K key instances selection

The data is structured in the following format: [score, coords, features]. In this format, 'score' represents the probability of being positive. 'Coords' represent the x and y coordinates, while 'features' represent the features extracted by a pretrained ResNet18 model.

python knn_feature.py --pt-path '../pt_files/' --save-path '../pt_files_knn/' --radius 50 --min-size 200 --ratio 0.4

We offer two implementation versions. If we opt for the iterative approach to obtain the super patch, using non-overlapping super patches can yield better performance. However, this may result in significantly higher CPU computational costs.

Transformer Aware of Instance Order and Distribution in MIL

for((FOLD=0;FOLD<4;FOLD++));
do
    python train.py --stage='train'\
    --config='Camelyon/TODMIL_995.yaml'  --gpus=0 --fold=$FOLD
    python train.py --stage='test'\
    --config='Camelyon/TODMIL_995.yaml'  --gpus=0 --fold=$FOLD
done
python metrics.py --config='Camelyon/TODMIL_995.yaml'