PGT

Prompt-based Grouping Transformer for Nucleus Detection and Classification

Overall Framework

Requisities

-python=3.8

-pytorch=1.12.1+cu102

Installation

Install mmcv using mim

pip install -U openmim
mim install mmcv-full==1.6.1

Git clone acformer

git clone https://github.com/LL3RD/PGT.git

Install

cd PGT
cd thirdparty/mmdetection
python -m pip install -e .
cd ../..
python -m pip install -e .

Dataset & Result

CoNSeP Dataset

Your can download 20x CoNSeP Dataset and Checkpoint from here.

Method	F1d	F1c	Model Weights	Config Files
PGT	0.738	0.613	Checkpoint	Config

BRCA Dataset

Your can download BRCA Dataset and Checkpoint from here.

Method	F1d	F1c	Model Weights	Config Files
PGT	0.799	0.523	Checkpoint	Config

Lizard Dataset

Your can download Original Lizard from the official website or Preprocessed Lizard and Checkpoint that is converted to hovernet consep format and split into patches.

Method	F1d	F1c	Model Weights	Config Files
PGT	0.779	0.538	Checkpoint	Config

Training

Modify your dataset path in configs and run

CUDA_VISIBLE_DEVICES=0 bash tools/dist_train.sh configs/Swin_CoNSeP_PGT.py 1 --work-dir=Path_to_your_work_dir

Evaluation

Modify your dataset path and checkpoint path in tools/inference and run

python tools/inference/inference_CoNSeP_group.py

python tools/inference/inference_BRCA_group.py

python tools/inference/inference_Lizard_group.py

Acknowledgement

• PGT is built based on SoftTeacher and MMDetection.