/VLPMNuD_part2

Official implementation of Zero-shot Nuclei Detection via Visual-Language Pre-trained Models. Experiment part 2.

Primary LanguagePython

Zero-shot Nuclei Detection via Visual-Language Pre-trained Models

Official implementation of Zero-shot Nuclei Detection via Visual-Language Pre-trained Models. The original paper link is here: arXiv link, [MICCAI link](to be update). The proposed method has two process steps:

  1. Generating raw GLIP prediction results. Code for this step is here.
  2. Self-training via YOLOX. This repository only provides source code for the second step.

Installation

  1. Create a Python environment based on the requirements.txt file.

  2. Build up YOLOX. Our project is developed based on YOLOX. Open the main directory and run:

python setup.py develop

Data Prepare

In the path datasets/COCO/annotations/, we placed some example raw GLIP prediction result (.json format) for quickly start. We recommend to use instances_train2017_best_raw_glip.json among thses files to reproduce our experiment result.

Self-training via YOLOX

Step 1: Train a new student via YOLOX:

python tools/train.py -f exps/default/yolox_s.py -d 1 -b 1 --fp16

Step 2: After that, if you want to repeat the self-training , take current student as a teacher and train a better student. You need to first employ the current student to prepare the pseudo-label, run:

python eval_and_regenerate_trainset.py -f exps/default/yolox_s.py -c "YOLOX_outputs/yolox_s/best_ckpt.pth" -b 1 -d 1 --conf 0.001 --fp16 --fuse

A new instance_train2017.json generated from the last student will replace the original label file at the path datasets/COCO/annotations/. It means now the last student will perform as a teacher, and provide its "pseudo label" to the next coming student. It is time to run Step 1 again.

You can repeat Step 1 and 2 until you meet satisfactory result. Normally, repeating 1 ~ 4 times should be sufficient. Sometimes the first student learns the best, but he teaches bad.

Citation

If you use VLPM in your work or wish to refer to the results published in this repo, please cite our paper:

@inproceedings{wu2023zero,
  title={Zero-shot Nuclei Detection via Visual-Language Pre-trained Models},
  author={Wu, Yongjian and Zhou, Yang and Saiyin, Jiya and Wei, Bingzheng and Lai, Maode and Shou, Jianzhong and Fan, Yubo and Xu, Yan},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={},
  year={2023},
  organization={Springer}
}