Pytorch implementation for the prompt tuning framework for MIL schemes described in the paper Prompt-MIL: Boosting Multi-Instance Learning Schemes via Task-specific Prompt Tuning, arxiv and video (MICCAI 2023, accepted for oral presentation).
Install Anaconda/miniconda.
Required packages:
$ conda create --name promptmil anaconda
$ conda activate promptmil
$ conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
$ pip install opencv-contrib-python kornia pytorch-lightning=2.0.2 torchmetrics=0.11.4 albumentations tqdm pandas wandb timm
You can also use docker or singularity. We provide the Dockerfile we used in our experiments. If you do not want to use wandb, you can change the logger in main.py.
Our dataset is organized as csv indicated datasets. All the images should be stored in a single directory and the path of this directory (dataset_root) should be passed to main.py. The labels and train/validation/test separation should be listed in a csv file. This csv file should contain 4 columns: wsi_id, label, len_img and fold. The csv files are provided.
slide_idis the filename of input image, not including file extension.labelis the integer label of an input WSI.len_imgrepresents the number of patches that this WSI contains. To reduce the time of dataloading thousands of patches, we merge every 200 patches vertically into a single patch (in merge_img_wsi.py). Each image should have the filenameslide_id_%d.jpg, where %d is the index 0,1,2,...foldrepresents "train" (fold > 0), "valid" (fold = 0) or "test" (fold < 0).
We used main.py to train and evaluate our framework.
usage: main.py [-h] [--output-dir OUTPUT_DIR] [--val_fold VAL_FOLD] [--batch-size-train BATCH_SIZE_TRAIN] [--batch-size-eval BATCH_SIZE_EVAL] [--dataset-name DATASET_NAME]
[--data-norm] [--data-mean DATA_MEAN] [--data-std DATA_STD] [--num-workers NUM_WORKERS] [--model MODEL] [--transfer-type TRANSFER_TYPE] [--network NETWORK]
[--num-prompt-tokens NUM_PROMPT_TOKENS] [--deep-prompt] [--prompt-dropout PROMPT_DROPOUT] [--project-prompt-dim PROJECT_PROMPT_DIM] [--epochs EPOCHS]
[--precision PRECISION] [--weight-decay WEIGHT_DECAY] [--lr LR] [--lr-factor LR_FACTOR] [--adam] [--loss-weight LOSS_WEIGHT] [--auto-loss-weight]
[--accumulate-grad-batches ACCUMULATE_GRAD_BATCHES] [--dropout-inst DROPOUT_INST] [--dropout-att DROPOUT_ATT] [--pretrained]
[--load-backbone-weight LOAD_BACKBONE_WEIGHT] [--load-weights LOAD_WEIGHTS] [--gpu-id GPU_ID] [--run-name RUN_NAME] [--tag TAG] [--clam-size CLAM_SIZE]
dataset_root dataset_csv
Useful arguments:
dataset_root # Path to all the patches in the dataset
dataset_csv # Path to the dataset csv file
--output-dir # Path of directory to store outputs
[--dataset-name] # Name of the dataset [tcga-brca|tcga-crc-cings|bright-3]
[--model] # Type of MIL schemes [dsmil|dsmil_ce|clam_sb|clam_mb|abmil|gabmil]
[--network] # Type of backbone network, e.g. dion_vit_tiny_patch16_192
[--transfer-type] # Type of fine-tuning, [prompt|end2end]
[--num-prompt-tokens] # Number of prompt tokens to add
[--num_epochs] # Number of training epochs
[--lr] # Initial learning rate
[--weight-decay] # Weight decay of the optimizer
[--batch-size-train] # The patch level batch size in training as shown in Figure 1(a)
[--batch-size-eval] # The patch level batch size in evaluation.
[--load-backbone-weight] # Path to the pretrained weights
[--run-name] # Project name
[--tag] # Experiment tag
[--gpu-id] # The id of one or multiple gpus to use, eg "0" or "0,1,2,3"
Note that the slide level batch size is always 1 as the input slide usually have different number of patches. We accumulate gradient of several batches by setting --accumulate-grad-batches, it is equivalent to a larger batch size.
The entire training can be done by, for example:
$ python main.py /path/to/merged/images dataset_csv/tcga_brca_label_cv.csv --output-dir /path/to/output --dataset-name tcga-brca --model dsmil --network dino_vit_tiny_patch16_192 --transfer-type "prompt" --epochs 50 --batch-size-train 512 --batch-size-eval 1024 --accumulate-grad-batches 8 --num-prompt-tokens 1 --lr 1e-4 --weight-decay 1e-2 --load-backbone-weight /path/to/the/weights/tcga-brca-vitt.pth --run-name PromptMIL --tag tcga_brca_prompt_vitt --gpu-id 0
$ python main.py /path/to/merged/images dataset_csv/bright3_label_cv.csv --output-dir /path/to/output --dataset-name bright-3 --model dsmil_ce --network dino_vit_tiny_patch16_192 --transfer-type "prompt" --epochs 50 --batch-size-train 512 --batch-size-eval 1024 --accumulate-grad-batches 8 --num-prompt-tokens 1 --lr 1e-4 --adam --weight-decay 5e-2 --load-backbone-weight /path/to/the/weights/bright-vitt.pth --run-name PromptMIL --tag bright3_prompt_vitt --gpu-id 1
$ python main.py /path/to/merged/images dataset_csv/tcga_crc_cings_label_cv.csv --output-dir /path/to/output --dataset-name tcga-crc-cings --model dsmil --network dino_vit_tiny_patch16_192 --transfer-type "prompt" --epochs 40 --batch-size-train 256 --batch-size-eval 512 --accumulate-grad-batches 8 --num-prompt-tokens 1 --lr 1e-4 --weight-decay 1e-2 --load-backbone-weight /path/to/the/weights/tcga-crc-vitt-cings.pth --run-name PromptMIL --tag tcga_crc_cings_prompt_vitt --gpu-id 2
Our DINO pretrained weights can be found here.
If you have any questions or concerns, feel free to directly contact us at Jingwei Zhang [email protected] or report issues.
Our framework used Pytorch Lightning. Thanks for this simple and efficient framework which facilitated our development. Part of our code was borrowed from DSMIL, CLAM, HIPT, and TransPath. Thanks for their outstanding work.
If you use the code or results in your research, please use the following BibTeX entry.
@inproceedings{zhang2023prompt,
title={Prompt-MIL: Boosting Multi-Instance Learning Schemes via Task-specific Prompt Tuning},
author={Zhang, Jingwei and Kapse, Saarthak and Ma, Ke and Prasanna, Prateek and Saltz, Joel and Vakalopoulou, Maria and Samaras, Dimitris},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
year={2023},
publisher="Springer Nature Switzerland",
pages="624--634",
}