Segmentation.md

PanDerm - Skin Lesion Segmentation

This implementation uses PanDerm for skin lesion segmentation.

Install Environment

Main dependencies: Pytorch, Pytorch Lightning, MMSegmentation

cd segmentation
pip install -r requirements.txt

Data Preparation

1. Download datasets:
   • ISIC2018 task1
   • HAM10000
2. Update _load_name() in ./datasets/dataset_seg.py with your data path

Fine-tuning with PanDerm Weight

1. Download Pretrained Weights

   • PanDerm weights

2. Set Pretrained Path

   • Update cae_weight in ./models/cae_seg.py
   • Modify pretrained parameter in run.sh

3. Start Training

   cd segmentation
   bash run.sh

   Note: Adjust the path config to your desired storage location

4. Evaluation

   • Add --evaluate to run.sh for evaluation mode
   • Model fine-tuned weights for evaluation on HAM10000 dataset Download here

   cd segmentation
   bash run.sh --evaluate

   This loads the checkpoint from your specified model storage path

Starter Code for Your Own Dermoscopic Images

See evaluate.ipynb to learn:

• Loading our pre-trained model
• Using it for skin lesion segmentation