This is code for paper “MBDS-Net: A Multi-Scale Boundary-Enhanced Denoising Diffusion Network for Medical Image Segmentation”

1.Prepare data

Download REFUGE dataset from https://refuge.grand-challenge.org/. Your dataset folder under "data" should be like:

/data/REFUGE/
├── train/
│   ├── imgs/
│   │   ├── case1.png
│   │   ├── case2.png
│   │   └── ...
│   └── masks/
│       ├── case1_mask.png
│       ├── case2_mask.png
│       └── ...
├── val/
│   ├── imgs/
│   │   ├── case1.png
│   │   ├── case2.png
│   │   └── ...
│   └── masks/
│       ├── case1_mask.png
│       ├── case2_mask.png
│       └── ...
└── test/
    ├── imgs/
    │   ├── case1.png
    │   ├── case2.png
    │   └── ...
    └── masks/
        ├── case1_mask.png
        ├── case2_mask.png
        └── ...

2.Data augmentation parameters

• Histogram Equalization

  • Applied to: Input RGB images
  • Method: ImageOps.equalize applied separately on R, G, B channels
  • Purpose: Enhance contrast before further processing.

• Resizing

  • Applied to: Both images and masks
  • Size: 256 × 256
  • Purpose: Standardize input size for training and evaluation.

• Random Horizontal Flip

  • Applied during: Training (if transform=True)
  • Probability: 0.5
  • Affects: Both image and mask simultaneously
  • Implementation: data_transforms.RandomHorizontalFlip(0.5)

• Normalization

  • Applied to: Input images only
  • Mean: [0.485, 0.456, 0.406]
  • Std: [0.229, 0.224, 0.225]
  • Performed after: ToTensor()

• Mask Preprocessing

  • Conversion: 0 (optic cup) → 1, 255 (background) → 0
  • Type: Single-channel float tensor
  • Purpose: Prepare binary mask for segmentation.

3.Training

python scripts/segmentation_train_my.py \
  --data_dir /data/REFUGE/train \
  --val_dir /data/REFUGE/val \
  --out_dir /model_save/REFUGE/ \
  --batch_size 8 \
  --lr 1e-4 \
  --use_fp16 False \
  --gpu_dev 0 \
  --schedule_sampler uniform \
  --save_interval 100 \
  --log_interval 100

4.Testing

python scripts/segmentation_sample_my.py \
  --data_dir /data/REFUGE/test \
  --batch_size 8 \
  --model_path /model_save/REFUGE/xxxx.pth \
  --num_ensemble 5 \
  --gpu_dev 0 \
  --out_dir /test_result/REFUGE/ \

4.Reference

5.Cite