DFU Segmentation Network

Deep learning model for Diabetic Foot Ulcer (DFU) segmentation using EfficientNet-B4 U-Net architecture, enhanced with a secondary verification classifier to reduce false positives.

File Structure

dfu-segmentation-net/
├── datasets/                        # Downloaded datasets (auto-generated)
├── inference_test/                  # Test images for inference
├── review/                          # Analysis and validation scripts
│   ├── fp.py                        # False positive analysis on non-wound datasets
│   ├── validate_skin_lesions.py     # Validation on skin lesion datasets
│   └── find_optimal_threshold.py    # Threshold optimization script
├── evaluate_pretrained_model.py     # Script to evaluate the segmentation model
├── pretrained.py                    # Segmentation model training and utils
├── template.py                      # Template/base code
├── wound_classifier.py              # Training script for the verification classifier
├── wound_classifier_inference.py    # Full pipeline inference (Segmentation + Verification)
├── pretrained_best_efficientnet_b4_unet_model.pth  # Segmentation model weights
└── wound_classifier_best_model.pth                 # Verification classifier weights

Features

1. Segmentation Model

The segmentation model performs binary semantic segmentation of Diabetic Foot Ulcers (DFU), producing precise pixel-level delineation of wound boundaries.

Architecture

• Encoder (Backbone): Uses EfficientNet-B4 pretrained on ImageNet as the feature extractor, capturing complex textures and patterns from input images.
• Decoder: Custom U-Net-style decoder that upsamples feature maps back to original resolution.
• Skip Connections: Connects corresponding layers between encoder and decoder (e.g., EfficientNet stage 6 to decoder layers), preserving fine-grained spatial details lost during downsampling—critical for accurate boundary detection.
• Input/Output: Takes RGB images (256×256) and outputs binary masks (1 for wound, 0 for background).

Training Configuration

• Loss Function: Dice Loss (optimizes overlap directly, better handling of class imbalance than Cross Entropy).
• Optimizer: Adam (learning rate: 0.001).
• Dataset: Trained on the "Wound Segmentation Dataset" from Kaggle.
• Training Strategy: Early stopping based on Dice score to prevent overfitting.

Performance

Based on evaluation on the test dataset:

• Dice Coefficient: ~89.3% (high overlap with ground truth).
• Intersection over Union (IoU): ~81.0%.
• Characteristics: Optimized for high Sensitivity (Recall) to minimize missed wounds, with the trade-off that false positives are filtered by the secondary Verification Classifier.

2. Verification Classifier (New)

• Goal: Filter out false positives (e.g., skin lesions, background noise) from the segmentation output.
• Architecture: Custom CNN taking 4-channel input (RGB Image + Predicted Mask) combined with segmentation percentage features.
• Training Data:
  • Positives: Wound Segmentation Dataset, Leprosy Chronic Wound Dataset.
  • Negatives: MiniImageNet (general objects), Skin Cancer & Skin Disease datasets (hard negatives).

Usage

1. Run Inference (Full Pipeline)

Run the complete pipeline (Segmentation -> Verification) on a folder of images:

py wound_classifier_inference.py --folder inference_test --output_json inference_results.json

Arguments:

• --folder: Input folder containing images.
• --seg_model: Path to segmentation model weights.
• --verif_model: Path to verification model weights.
• --output_json: Path to save results.

2. Train Verification Classifier

Train the binary classifier to distinguish between true wounds and false positives:

py wound_classifier.py

This script handles dataset downloading, balancing, and training automatically.

3. Analyze False Positives

Evaluate the segmentation model's performance on non-wound datasets to measure false alarm rates:

py review/fp.py

Model Predictions

Metric	CNN (0.5)	ViT (0.5)	CNN (0.41)	ViT (0.70)
AZH Wound Dataset (Sensitivity Test)
Total Images	213	213	213	213
True Positives (TP)	121	132	137	125
False Negatives (FN)	92	81	76	88
Sensitivity (Recall)	56.81%	61.97%	64.32%	58.69%
Avg Confidence	0.5471	0.6056	0.5471	0.6056
Avg Segmentation %	4.51%	4.38%	4.51%	4.38%
Skin Lesion Dataset (Specificity Test)
Total Images	500	500	500	500
False Positives (FP)	69	57	91	32
True Negatives (TN)	431	443	409	468
Specificity	86.20%	88.60%	81.80%	93.60%
False Positive Rate	13.80%	11.40%	18.20%	6.40%
Avg Confidence	0.2142	0.1208	0.2142	0.1208
Avg Segmentation %	3.85%	3.72%	3.85%	3.72%
Combined Metrics
Precision	63.68%	69.84%	60.09%	79.62%
F1 Score	60.00%	65.67%	62.13%	67.57%

_{Note: Results shown for both default threshold (0.5) and optimized thresholds (CNN: 0.41, ViT: 0.70) determined by F1 score maximization. Bold values indicate best performance. Lower is better for False Negatives and False Positive Rate; higher is better for others.}