Awesome

TPRO: Text-prompting-based Weakly Supervised Histopathology Tissue Segmentation

This is the official pytorch implementation of our MICCAI 2023 paper "TPRO: Text-prompting-based Weakly Supervised Histopathology Tissue Segmentation".

Preparation

Download LUAD-HistoSeg and BCSS-WSSS datasets and orgnize the directory sctructure in the following format:

data/
|--LUAD-HistoSeg
   |--train
      |--img
   |--test
      |--img
      |--mask
   |--valid
      |--img
      |--mask
|--BCSS-WSSS
   |--train
      |--img
   |--test
      |--img
      |--mask
   |--valid
      |--img
      |--mask

The ImageNet-1k pre-trained weights of vision encoder can be download from the official SegFormer implementation.

Train a CLassification Network

# LUAD-HistoSeg
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 16732 train_cls.py --config ./work_dirs/luad/classification/config.yaml
# BCSS-WSSS
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 16372 train_cls.py --config ./work_dirs/bcss/classification/config.yaml

Extract Pseudo Labels

# LUAD-HistoSeg
python evaluate_cls.py --dataset luad --model_path path/to/classification/model --save_dir ./work_dirs/luad/classification/predictions --split train
# BCSS-WSSS
python evaluate_cls.py --dataset bcss --model_path path/to/classification/model --save_dir ./work_dirs/bcss/classification/predictions --split train

Train a Segmentation Network

# LUAD-HistoSeg
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 16372 train_seg.py --config ./work_dirs/luad/segmentation/config.yaml
# BCSS-WSSS
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 16372 train_seg.py --config ./work_dirs/bcss/segmentation/config.yaml