Skip to content

Latest commit

 

History

History
54 lines (41 loc) · 3.17 KB

README.md

File metadata and controls

54 lines (41 loc) · 3.17 KB
Raw

NPCNet

The official implementation of work entitled "NPCNet: Jointly Segment Primary Nasopharyngeal Carcinoma Tumor and Metastatic Lymph Nodes in MR Images", accepted by IEEE Transactions on Medical Imaging, 2022. The paper is now under publishing and now available at IEEE Early Access

Notice

How to train the model stably? While doing ablation study, we found that training the model with BEM from scratch is unstable, which means that the performance of the BEM depends on the low-resolution segmentation results. The way we train the model:

  1. Train the base model with PEM and SEM, you need to modify the model code to remove the BEM, and remove the "point_loss" in "train_npcnet.py" while calculating loss. Save the best model (calculate the seg metrics utilizing fine mask).
  2. Modify the model code by adding BEM in model code and adding point_loss in training code, load the weights from step 1.
  3. Train the whole model. The learning rate of the parameters in base model, PEM and SEM is 1/10 of those in BEM (You will find the optimizer lr settings and loss function settings in the codes).
  4. The trained model weights on our datasets will be uploaded soon.
  5. The datasets used in this paper will be made public available by the hospital mentioned in our paper. Please contact with the corresponding authors.

Method

Model Architecture

Training

Requirements

  • Python>=3.7
  • PyTorch ≥ 1.4 and torchvision that matches the PyTorch installation. You can install them together at pytorch.org to make sure of this
  • Other libs in the requirements.txt file

Prepare datasets

You should prepare your own dataset according to the dataloader.py file. The input MRI images are concatenated in channel (sequence is T1, T2, T1c) and saved in .npy format with data type of float32. The GT files (segmentation mask) should be saved in .npy format with data type of uint8.

Model weights

The weights trained on our dataset are available at Google Cloud.

Training details

To achieve the best results and avoid unstable training, you can train the model according to these steps:

  1. Train a model without the proposed BEM (which is backbone + PEM + SEM) and save the weights.
  2. Modify the model definition code by adding the proposed BEM (which is NPCNet).
  3. Train the whole NPCNet with the pretrained weights in step 1 (parameters in the BEM are initialized randomly). The learning rate for backbone, PEM, and SEM is 1e-3 and the learning rate for BEM is 1e-2.

Q&A

Leave your question in the issue of this repo. We would answer your question ASAP.

Citation

@ARTICLE{9684475,
  author={Li, Yang and Dan, Tingting and Li, Haojiang and Chen, Jiazhou and Peng, Hong and Liu, Lizhi and Cai, Hongmin},
  journal={IEEE Transactions on Medical Imaging},
  title={NPCNet: Jointly Segment Primary Nasopharyngeal Carcinoma Tumors and Metastatic Lymph Nodes in MR Images},
  year={2022},
  volume={41},
  pages={1639-1650},
  doi={10.1109/TMI.2022.3144274}
}