DiaFoot.AI 🦶

State-of-the-Art Diabetic Foot Ulcer Segmentation

Model Description

DiaFoot.AI is a deep learning model for automatic segmentation of diabetic foot ulcers from clinical images. It uses a U-Net++ architecture with an EfficientNet-B4 encoder pretrained on ImageNet.

Performance

Metric	DiaFoot.AI	FUSeg 2021 Winner	Improvement
IoU	85.58%	80.30%	+5.28%
Dice	92.23%	89.23%	+3.00%

Cross-Validation Results

Fold	IoU	Dice
Fold 1	83.22%	90.84%
Fold 2	85.81%	92.36%
Fold 3	82.16%	90.20%
Fold 4	85.00%	91.89%
Fold 5	84.44%	91.56%
Mean	84.13% ± 1.30%	91.37% ± 0.77%

Usage

Installation

pip install torch segmentation-models-pytorch albumentations huggingface_hub

Download and Load Model

from huggingface_hub import hf_hub_download
import torch
import segmentation_models_pytorch as smp

# Download model
model_path = hf_hub_download("RuthvikBandari/DiaFootAI", "best_model.pt")

# Create model architecture
model = smp.UnetPlusPlus(
    encoder_name="efficientnet-b4",
    encoder_weights=None,
    in_channels=3,
    classes=1
)

# Load weights
checkpoint = torch.load(model_path, map_location="cpu", weights_only=False)
state_dict = checkpoint["model_state_dict"]
# Remove 'model.' prefix if present
state_dict = {k.replace("model.", "", 1): v for k, v in state_dict.items()}
model.load_state_dict(state_dict)
model.eval()

Inference

import cv2
import numpy as np
import albumentations as A
from albumentations.pytorch import ToTensorV2

# Preprocessing
transform = A.Compose([
    A.LongestMaxSize(max_size=512),
    A.PadIfNeeded(min_height=512, min_width=512, border_mode=cv2.BORDER_CONSTANT),
    A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ToTensorV2()
])

# Load and preprocess image
image = cv2.imread("wound_image.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
transformed = transform(image=image)
input_tensor = transformed["image"].unsqueeze(0)

# Predict
with torch.no_grad():
    output = model(input_tensor)
    mask = torch.sigmoid(output).squeeze().numpy()
    binary_mask = (mask > 0.5).astype(np.uint8)

Model Architecture

Architecture: U-Net++
Encoder: EfficientNet-B4 (ImageNet pretrained)
Input Size: 512 × 512 × 3
Output: Single channel probability map
Parameters: 20.8M

Training Details

Loss: Focal Tversky + BCE with warmup
Optimizer: AdamW
Learning Rate: 3e-4 (decoder), 1e-5 (encoder)
Batch Size: 8
Epochs: 100
Augmentation: Heavy (flips, rotations, color jitter, elastic transform)
Test Time Augmentation: 8 transforms

Dataset

Trained on FUSeg 2021 Challenge Dataset:

810 training images
200 validation images
Pixel-wise annotations by wound care experts

Citation

@software{bandari2025diafootai,
  author = {Bandari, Ruthvik},
  title = {DiaFoot.AI: Deep Learning for Diabetic Foot Ulcer Segmentation},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/Ruthvik-Bandari/DiaFoot.AI},
  note = {Achieves 85.58% IoU, surpassing MICCAI FUSeg 2021 winner by +5.28%}
}

License

CC BY-NC 4.0 - You must give appropriate credit and may not use for commercial purposes without permission.

Author

Ruthvik Bandari
Master's in Applied Artificial Intelligence
Northeastern University

GitHub: @Ruthvik-Bandari
LinkedIn: Ruthvik Bandari

Disclaimer

This model is intended for research purposes only. Not for clinical diagnosis without professional medical oversight.

Downloads last month: 8