LViT

This repo is the official implementation of "LViT: Language meets Vision Transformer in Medical Image Segmentation" Paper Link

Requirements

Install from the requirements.txt using:

pip install -r requirements.txt

Questions about NumPy version conflict. The NumPy version we use is 1.17.5. We can install bert-embedding first, and install NumPy then.

Usage

1. Data Preparation

1.1. QaTa-COV19, MosMedData+ and MoNuSeg Datasets (demo dataset)

The original data can be downloaded in following links:

QaTa-COV19 Dataset - Link (Original)
MosMedData+ Dataset - Link (Original)
MoNuSeG Dataset (demo dataset) - Link (Original)

The text annotation of QaTa-COV19 has been released!

(Note: The text annotation of QaTa-COV19 train and val datasets download link. The partition of train set and val set of QaTa-COV19 dataset download link. The text annotation of QaTa-COV19 test dataset download link.)

(Note: The text annotation of MosMedData+ dataset will be released in the future.)

If you use the datasets provided by us, please cite the LViT.

1.2. Format Preparation

Then prepare the datasets in the following format for easy use of the code:

├── datasets
    ├── QaTa-Covid19
    │   ├── Test_Folder
    |   |   ├── Test_text.xlsx
    │   │   ├── img
    │   │   └── labelcol
    │   ├── Train_Folder
    |   |   ├── Train_text.xlsx
    │   │   ├── img
    │   │   └── labelcol
    │   └── Val_Folder
    |	    ├── Val_text.xlsx
    │       ├── img
    │       └── labelcol
    └── MosMedDataPlus
        ├── Test_Folder
        |   ├── Test_text.xlsx
        │   ├── img
        │   └── labelcol
        ├── Train_Folder
        |   ├── Train_text.xlsx
        │   ├── img
        │   └── labelcol
        └── Val_Folder
            ├── Val_text.xlsx
            ├── img
            └── labelcol

2. Training

2.1. Pre-training

You can replace LVIT with U-Net for pre training and run:

python train_model.py

2.2. Training

You can train to get your own model. It should be noted that using the pre-trained model in the step 2.1 will get better performance or you can simply change the model_name from LViT to LViT_pretrain in config.

python train_model.py

3. Evaluation

3.1. Get Pre-trained Models

Here, we provide pre-trained weights on QaTa-COV19 and MoNuSeg, if you do not want to train the models by yourself, you can download them in the following links:

(Note: the pre-trained model will be released in the future.)

3.2. Test the Model and Visualize the Segmentation Results

First, change the session name in Config.py as the training phase. Then run:

python test_model.py

You can get the Dice and IoU scores and the visualization results.

4. Results

Dataset	Model Name	Dice (%)	IoU (%)
QaTa-COV19	U-Net	79.02	69.46
QaTa-COV19	LViT-T	83.66	75.11
MosMedData+	U-Net	64.60	50.73
MosMedData+	LViT-T	74.57	61.33
MoNuSeg	U-Net	76.45	62.86
MoNuSeg	LViT-T	80.36	67.31
MoNuSeg	LViT-T w/o pretrain	79.98	66.83

4.1. More Results on other datasets

Dataset	Model Name	Dice (%)	IoU (%)
BKAI-Poly	LViT-TW	92.07	80.93
ESO-CT	LViT-TW	68.27	57.02

5. Reproducibility

In our code, we carefully set the random seed and set cudnn as 'deterministic' mode to eliminate the randomness. However, there still exsist some factors which may cause different training results, e.g., the cuda version, GPU types, the number of GPUs and etc. The GPU used in our experiments is 2-card NVIDIA V100 (32G) and the cuda version is 11.2. And the upsampling operation has big problems with randomness for multi-GPU cases. See https://pytorch.org/docs/stable/notes/randomness.html for more details.

Reference

Citation

@article{Li2022LViT,
  title={LViT: Language meets Vision Transformer in Medical Image Segmentation},
  author = {Li, Zihan and Li, Yunxiang and Li, Qingde and Wang, Puyang and Zhang, You and Guo, Dazhou and Lu, Le and Jin, Dakai and Hong, Qingqi},
  journal={arXiv preprint arXiv:2206.14718},
  year={2022}
}

JAYCHOU2020/LViT