FPT: PETL for High-resolution Medical Image Classification

This is the official implementation of the papers:

Y. Huang, P. Cheng, R. Tam, and X. Tang, "Boosting Memory Efficiency in Transfer Learning for High-Resolution Medical Image Classification", IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2025. [tnnls] [arXiv]

Y. Huang, P. Cheng, R. Tam, and X. Tang, "Fine-grained Prompt Tuning: A Parameter and Memory Efficient Transfer Learning Method for High-resolution Medical Image Classification", In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2024. [miccai] [arXiv]

We present Fine-grained Prompt Tuning (FPT) and FPT+ for medical image classification. FPT and FPT+ are parameter-efficient transfer learning (PETL) methods that significantly improve memory efficiency over existing PETL methods, especially in the high-resolution context commonly encountered in medical image analysis.

Performance

Performance using a ViT-B backbone initialized with ImageNet-21K pre-trained weights:

Method	# Learnable Parameters	Memory Usage	Average AUC
Full fine-tuning	100	23128	88.82
Prompt-tuning	0.17	20582	83.68
Adapter	2.03	19360	84.14
LoRA	0.68	20970	85.94
FPT	1.81	1824	86.40
FPT+	1.03	736	87.12

Installation

To install the dependencies, run:

git clone https://github.com/YijinHuang/FPT
conda create -n fpt python=3.8
conda activate fpt
pip install -r requirements.txt

Dataset

Eight publicly accessible datasets are used in this work:

• Messidor-2 (Fundus image) [Images] [Labels]
• DDR (Fundus image) [Homepage]
• ISIC 2016 (Dermoscopic image) [Homepage]
• ISIC 2018 (Dermoscopic image) [Homepage]
• Mini-DDSM (Mammography) [Homepage]
• CMMD (Mammography) [Homepage]
• COVID (Chest X-ray) [Homepage]
• CHNCXR (Chest X-ray) [Homepage]

How to Use

We use the Messidor-2 dataset as an example in the instructions.

1. Build dataset

Organize the Messidor-2 dataset as follows:

messidor2_dataset/
├── train/
│   ├── class1/
│   │   ├── image1.jpg
│   │   ├── image2.jpg
│   ├── class2/
│   │   ├── image3.jpg
│   │   ├── image4.jpg
│   ├── class3/
│   ├── ...
├── val/
├── test/

Ensure the val and test directories have the same structure as train. Then, update the data_path value in /configs/dataset/messidor2.yaml with the path to the Messidor-2 dataset.

2. Preloading

Pre-store features from LPM by running:

python preload.py dataset=messidor2

3. Training

To train the model, run:

python main.py dataset=messidor2

Train on Your Own Dataset

1. Build your dataset

Organize your dataset similarly to Messidor-2.

2. Create and update configurations

Update the configurations marked as '???' in /configs/dataset/customized.yaml.

3. Preloading

python preload.py dataset=customized

If the image resolution is very large and causes out-of-memory issues, decrease the batch size for preloading:

python preload.py dataset=customized ++train.batch_size=4

After preloading, FPT or FPT+ can still be run with a large batch size.

4. Training

To train the model, run:

python main.py dataset=customized

Other Configurations

You can update the configurations in /configs. Hydra is employed to manage configurations. For advanced usage, please check the Hydra documentation.

1. Run FPT

The default method is FPT+. To run FPT, update the command with network=FPT:

python preload.py dataset=messidor2 network=FPT
python main.py dataset=messidor2 network=FPT

Note that FPT+ and FPT doesn't share the preloaded features.

2. Pre-trained model

Most ViT-based models from Hugging Face uploaded by google/facebook/timm can be directly employed. Default pre-trained weights is google/vit-base-patch16-384. To change the LPM, set the pre-trained path in /configs/network/FPT+.yaml or update the command to:

python main.py dataset=messidor2 ++network.pretrained_path=google/vit-base-patch16-384

Validated pre-trained weights in this work:

• google/vit-base-patch16-384
• google/vit-large-patch16-384
• facebook/dino-vitb8
• facebook/dino-vitb16

3. Disable prelading

To disable preloading, set the 'preload_path' in /configs/dataset/your_dataset.yaml to 'null' or update the command to:

python main.py dataset=messidor2 ++dataset.preload_path=null

4. Learning rate

To change the learning rate, set the 'learning_rate' in /configs/dataset/your_dataset.yaml or update the command to:

python main.py dataset=messidor2 ++dataset.learning_rate=0.0001

5. Random seed

To control randomness, set the 'seed' to a non-negative integer in /configs/config.yaml or update the command to:

python main.py dataset=messidor2 ++base.seed=0

Citation

If you find this repository useful, please cite the papers:

@article{huang2025fptp,
  author={Huang, Yijin and Cheng, Pujin and Tam, Roger and Tang, Xiaoying},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={Boosting Memory Efficiency in Transfer Learning for High-Resolution Medical Image Classification}, 
  year={2025},
  volume={36},
  number={9},
  pages={17280-17294},
  keywords={Adaptation models;Memory management;Transfer learning;Training;Tuning;Medical diagnostic imaging;Graphics processing units;Transformers;Image classification;Costs;High-resolution medical image classification;large-scale pretrained models;memory-efficient transfer learning;parameter-efficient transfer learning (PETL)},
  doi={10.1109/TNNLS.2025.3569797}}
}

@inproceedings{huang2024fine,
  title={Fine-grained prompt tuning: A parameter and memory efficient transfer learning method for high-resolution medical image classification},
  author={Huang, Yijin and Cheng, Pujin and Tam, Roger and Tang, Xiaoying},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={120--130},
  year={2024},
  organization={Springer}
}