SEG-PEFT is a repository for medical image segmentation research, focused on polyp detection in endoscopic images. The project explores the effectiveness of LoRA (Low-Rank Adaptation) for medical image segmentation tasks, specifically comparing full fine-tuning vs. parameter-efficient fine-tuning approaches on the Kvasir-SEG dataset.

You shall download the dataset at the following link https://datasets.simula.no/downloads/kvasir-seg.zip

The Kvasir-SEG dataset is a medical image dataset specifically designed for polyp segmentation in endoscopic images. The dataset contains:

• Images: Endoscopic images of the GI tract containing polyps
• Masks: Binary segmentation masks where the polyp regions are marked
• Size: 1000 image-mask pairs for training and validation
• Task: Semantic segmentation to identify and segment polyp regions

To use the dataset with this project, you must place it in the following directory structure:

SEG-PEFT/
├── data/
│   └── Kvasir-SEG/
│       ├── images/
│       │   ├── cju0bygl98aar07367zog10cj.jpg
│       │   ├── cju0byglb8abs073685yd3kjn.jpg
│       │   └── ... (998 more image files)
│       └── masks/
│           ├── cju0bygl98aar07367zog10cj.jpg
│           ├── cju0byglb8abs073685yd3kjn.jpg
│           └── ... (998 more mask files)

The dataset is organized in two directories:

• data/Kvasir-SEG/images/: Contains the endoscopic images in RGB format
• data/Kvasir-SEG/masks/: Contains the corresponding binary masks where polyps are labeled as white (255) and background as black (0)

During training, the images are preprocessed with the following augmentations:

• Resize to 512x512 pixels
• Horizontal flipping (50% probability)
• Random brightness and contrast adjustments (20% probability)

Forst of all you must execute the following commmands to clone and install the dependencies:

git clone https://github.com/rossoc/SEG-PEFT
cd SEG-PEFT
pip install -r requirements

Furthermore, you may want to activate the environment:

source .venv/bin/activate

To train the SegFormer model without LoRA (full fine-tuning):

python scripts/train_segformer.py --epochs 30 --lr 5e-5 --save-dir my_experiment

Additional options:

• --epochs: Number of training epochs (default: 30)
• --lr: Learning rate (default: 5e-5)
• --save-dir: Directory name to save the model (required)
• --seed: Random seed for reproducibility (default: 42)

To train the SegFormer model with LoRA (parameter-efficient fine-tuning):

python scripts/train_segformer_lora.py --epochs 30 --lr 5e-5 --rank 8 --lora-alpha 32 --lora-dropout 0.1 --save-dir my_lora_experiment

Additional options:

• --epochs: Number of training epochs (default: 30)
• --lr: Learning rate (default: 5e-5)
• --rank: LoRA rank (default: 8)
• --lora-alpha: LoRA alpha parameter (default: 32)
• --lora-dropout: LoRA dropout rate (default: 0.1)
• --save-dir: Directory name to save the model (required)
• --seed: Random seed for reproducibility (default: 42)