LangMamba

Official implementation of "LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models"

Updates

March, 2025: initial commit.

Data Preparation

The 2016 AAPM-Mayo dataset can be downloaded from: CT Clinical Innovation Center (B30 kernel)
The 2020 AAPM-Mayo dataset can be downloaded from: cancer imaging archive

Dataset structre:

Mayo2016_2d/
  |--train/
      |--quarter_1mm/
        train_quarter_00001.npy
        train_quarter_00002.npy
        train_quarter_00003.npy
        ...
      |--full_1mm/
        train_full_00001.npy
        train_full_00002.npy
        train_full_00003.npy
        ...
  |--test/
      |--quarter_1mm
      |--full_1mm

Requirements

- Linux Platform
- torch==1.12.1+cu113 # depends on the CUDA version of your machine
- torchvision==0.13.1+cu113
- Python==3.8.0
- numpy==1.22.3

Traning and & Inference

Training of LLM-guided NDCT autoencoder:

Build per-layer candidate token:python process_words.py
Then we used the official repository of VQ-GAN (https://github.com/CompVis/taming-transformers) to set up training. Please refer to (models/taming) to learn about our modifications to original VQ-GAN.

Employment of LEDA for training the denoising model:

python train.py  --name LangMamba(experiment_name)   --model langmamba  --dataroot /data/zhchen/Mayo2016_2d(path to images) --lr 0.0002 --gpu_ids 6,7 --print_freq 25 --batch_size 8 --lr_policy cosine

Inference & testing:

python test.py  --name LangMamba(experiment_name)   --model langmamba --results_dir test_results --result_name test_results(path to save image)   --gpu_ids 6 --batch_size 1 --eval

Please refer to options files for more setting.