RuthvikBandari/DiaFoot.AI-v2

DiaFoot.AI v2 — Diabetic Foot Ulcer Detection & Segmentation

⚠️ IMPORTANT DISCLAIMER: This is an academic research project developed for educational purposes as part of the AAI6620 Computer Vision course at Northeastern University. This software is NOT a medical device, is NOT FDA-cleared, and is NOT intended for clinical use, diagnosis, treatment, or any medical decision-making. It does not replace professional medical judgment. Always consult a qualified healthcare provider for any medical concerns. The authors assume no liability for any use of this software in clinical or diagnostic settings.

---

A production-grade multi-task pipeline for automated diabetic foot ulcer (DFU) detection, wound boundary segmentation, and clinical wound assessment. Built to demonstrate modern computer vision techniques applied to medical imaging.

---

Clinical Motivation

Diabetic foot ulcers affect 15–25% of diabetic patients in their lifetime, with 85% of diabetes-related amputations preceded by a foot ulcer. Early detection and accurate wound measurement can reduce amputation rates by up to 85%. DiaFoot.AI explores how deep learning can automate wound boundary detection to potentially support clinical workflows in the future.

Why v2: Lessons from v1

The original DiaFoot.AI (v1) achieved 84.93% IoU and 91.73% Dice — numbers that looked strong but masked two fundamental flaws:

1. Training data contained only ulcer images. The model never learned what healthy skin looks like, so it predicted ulcers on every input — zero clinical specificity. A model that calls everything a wound is clinically useless.
2. No data cleaning pipeline. Raw scraped images were fed directly into training with no quality audit, duplicate detection, or label verification.

DiaFoot.AI v2 is a ground-up rebuild that fixes both problems through a multi-task cascaded pipeline and rigorous data engineering.

---

Architecture

The system uses a cascaded pipeline (Strategy A), validated by ablation to outperform joint multi-task training:

Input Image
    │
    ▼
┌───────────────────────────┐
│  Triage Classifier        │
│  EfficientNet-V2-M        │
│                           │
│  → Healthy                │  ← Stop. No wound detected.
│  → Non-DFU Condition      │  ← Stop. Not a diabetic ulcer.
│  → DFU Detected           │  ← Proceed to segmentation.
└───────────┬───────────────┘
            │ (DFU only)
            ▼
┌───────────────────────────┐
│  Wound Segmenter          │
│  U-Net++ / EfficientNet-B4│
│  + scSE Attention         │
│                           │
│  → Pixel-wise wound mask  │
│  → Wound area (mm²)       │
│  → Boundary metrics       │
└───────────────────────────┘

Why cascaded? The data composition ablation proved that the segmenter performs best when trained exclusively on DFU images (85.89% Dice). Adding non-DFU wounds actually hurt performance (68.71% Dice) because the model gets confused learning two different wound morphologies simultaneously. The classifier handles triage; the segmenter focuses on what it does best.

---

Results

5-Fold Cross-Validated Segmentation (DFU)

The primary result, validated across 5 independent train/val splits for statistical rigor:

Fold	Dice	IoU
0	84.69%	78.03%
1	86.10%	79.87%
2	85.98%	79.00%
3	84.74%	78.07%
4	85.66%	78.54%
Mean ± Std	85.43 ± 0.61%	78.70 ± 0.68%

The standard deviation of ±0.61% confirms the model performs consistently regardless of data partitioning.

Test Set Evaluation (DFU, n=285)

Metric	Value	Clinical Interpretation
Dice Score	85.89%	Strong pixel-level wound overlap
IoU (Jaccard)	79.35%	Solid intersection accuracy
HD95	17.3 px	95th percentile boundary distance
NSD@2mm	85.86%	86% of predicted boundary within 2mm of ground truth
NSD@5mm	94.74%	95% within 5mm — clinically excellent
Wound Area Error	1.1%	Predicted 1,342 mm² vs ground truth 1,358 mm²

Data Composition Ablation

The single most important experiment — proving that data composition matters more than architecture:

Training Data	Best Dice	Val Loss	Overfitting Ratio
DFU-only (1,881 images)	87.44%	0.1078	1.3x (minimal)
DFU-only (1,010 images)	85.27%	0.1057	1.0x (none)
DFU + non-DFU	68.71%	0.4187	1.4x
All classes (DFU + healthy + non-DFU)	84.14%*	0.6723	2.9x (heavy)

*Inflated by healthy images scoring perfectly on empty masks.

Key finding: Adding 871 more DFU images (AZH wound care center data) improved Dice by +2.17% with no other changes. Data quality and quantity matter more than architecture complexity.

Architecture Comparison

Model	Best Dice	Parameters	Notes
U-Net++ / EfficientNet-B4 + scSE	87.44%	~25M	Best performance
FUSegNet / EfficientNet-B7 + P-scSE	69.60%	~66M	Too many parameters for dataset size

Test-Time Augmentation (TTA)

Metric	Without TTA	With TTA (16-aug)	Improvement
Dice	57.38%*	61.26%*	+3.88%
IoU	52.28%*	56.29%*	+4.01%
HD95	87.47	84.88	-2.59 (better)

*Overall numbers including non-DFU images; DFU-specific TTA improvement follows the same trend.

Fairness Analysis (ITA-Stratified)

ITA Group	Count	Dice	IoU	HD95
Brown	285	85.89%	79.35%	17.3
Fairness gap	—	0.00%	—	—

Limitation: The dataset is predominantly composed of a single ITA skin tone group (Brown). While no fairness gap exists within the represented population, the model has not been validated across the full Fitzpatrick I–VI spectrum. ITA computation on wound images is confounded by wound bed color; a clinical deployment would require ITA measurement from non-wound skin regions specifically.

---

Dataset

Composition (8,105 total samples)

Category	Images	Purpose
DFU	2,119	Wound segmentation training (FUSeg + AZH)
Healthy Feet	3,300	True negatives for classifier
Non-DFU Conditions	2,686	Hard negatives (general wounds, not DFU)

Sources

Dataset	Images	Type
FUSeg 2021 (UWM BigData Lab)	1,010	DFU with segmentation masks
AZH Wound Care Center	1,109	Clinical wound patches with masks
Kaggle DFU Patches	543	Healthy foot skin patches
Mendeley Wound Dataset (Normal)	2,757	Healthy foot images
Mendeley Wound Dataset (Wounds)	2,686	Non-DFU wound images with masks

Data Pipeline

All images pass through a production cleaning pipeline:

1. Integrity check — verify every image opens and is not corrupt
2. Mask validation — binary format check, dimension alignment, coverage statistics
3. Deduplication — perceptual hash (dHash) to remove cross-dataset duplicates
4. Preprocessing — resize to 512×512 (aspect-preserving pad), CLAHE contrast enhancement, mask binarization
5. Stratified splits — 70/15/15 train/val/test, stratified by class and ITA skin tone group, zero data leakage verified

---

Tech Stack

Component	Library	Version
Deep Learning	PyTorch	2.10.0
Medical Imaging	MONAI	1.5.2
Segmentation	Segmentation Models PyTorch	0.5.0
Augmentation	Albumentations	1.4.24
Data Quality	CleanVision, Cleanlab	Latest
API	FastAPI	0.133.0
Inference	ONNX Runtime	1.24.2
Linting	Ruff	0.15.2
Compute	Northeastern Explorer HPC (H200/A100 GPUs)	—

---

Quick Start

Installation

git clone https://github.com/Ruthvik-Bandari/DiaFoot.AI.git
cd DiaFoot.AI
pip install -r requirements.txt

Inference (Single Image)

python scripts/predict.py --image path/to/foot.jpg --device cuda

Training

# Train classifier (all 3 classes)
python scripts/train.py --task classify --epochs 50 --device cuda

# Train segmenter (DFU-only — proven best by ablation)
python scripts/run_ablation.py --variant dfu_only --epochs 100 --device cuda

5-Fold Cross-Validation

# Submit as SLURM array job (parallel)
sbatch slurm/run_cv.sh

# Or run individual folds
python scripts/run_cross_val.py --fold 0 --device cuda --epochs 100
python scripts/run_cross_val.py --fold 1 --device cuda --epochs 100
# ... folds 2, 3, 4

Evaluation

python scripts/evaluate.py \
    --task segment \
    --checkpoint checkpoints/ablation_dfu_only/best_epoch090_0.1078.pt \
    --device cuda

ONNX Export

python scripts/export_onnx.py \
    --checkpoint checkpoints/ablation_dfu_only/best_epoch090_0.1078.pt \
    --output models/diafoot_segmenter.onnx \
    --validate --benchmark

FastAPI Server

uvicorn src.deploy.app:app --host 0.0.0.0 --port 8000
# POST /predict with image file

---

Project Structure

DiaFoot.AI/
├── configs/                    # YAML configs for training, models, data
├── data/
│   ├── raw/                    # Original datasets (DVC tracked)
│   ├── processed/              # Cleaned, preprocessed 512×512 images
│   ├── splits/                 # Train/val/test CSVs
│   └── metadata/               # Quality reports, ITA scores
├── src/
│   ├── data/                   # Dataset classes, augmentation, cleaning
│   ├── models/                 # U-Net++, FUSegNet, classifier, MedSAM2
│   ├── training/               # Trainer, losses, schedulers, EMA
│   ├── evaluation/             # Metrics, fairness, calibration, robustness
│   ├── inference/              # Pipeline, TTA, postprocessing
│   └── deploy/                 # FastAPI app
├── scripts/                    # Entry points for train, eval, export
├── slurm/                      # HPC job scripts
├── results/                    # Metrics, figures, reports
├── checkpoints/                # Trained model weights
└── tests/                      # Unit tests

---

Peer Feedback Integration

Every piece of peer feedback from the AAI6620 course review was mapped to a specific implementation:

Feedback	From	Implementation
How did augmentation handle skin tone diversity?	Sudeep K.S.	ITA-stratified fairness audit
Add attention mechanisms to reduce false positives	Shivam Dubey	scSE attention in U-Net++ decoder
Report performance relative to inter-annotator agreement	Yucheng Yan	Ceiling analysis framework
Tie uncertainty to clinical output	Yash Jain	TTA-based uncertainty maps
Prioritize ablation studies over deployment	Yucheng Yan	Data composition ablation as core experiment
Addressing algorithmic bias is a critical ethical hurdle	Ching-Yi Mao	ITA fairness audit with honest limitation disclosure

---

Honest Limitations

1. Classifier accuracy (100%) is a dataset artifact. The three data categories come from visually distinct sources (different cameras, backgrounds). The classifier learned "which dataset" rather than "which condition." A production system requires same-source data across all classes.

2. Wagner staging was not trained. The architecture supports it, but clinical grade labels were unavailable. This is acknowledged as future work requiring clinical partnerships.

3. Limited skin tone diversity. The dataset is predominantly a single ITA group. Fairness conclusions cannot be generalized to the full Fitzpatrick I–VI spectrum. A clinical system would require validation across diverse skin tones.

4. Only 2 of 5 architectures were fully trained. FUSegNet underperformed; MedSAM2 LoRA and nnU-Net v2 were implemented but not trained due to time constraints.

5. Not validated on standardized benchmarks. Results are on our own data splits. Comparison against the DFUC 2022 challenge leaderboard would require access to their test set.

---

Regulatory & Ethical Notice

This project is developed strictly for academic and educational purposes. It is part of the AAI6620 Computer Vision coursework at Northeastern University.

This software:

• Is NOT a medical device as defined by the FDA, EU MDR, or any regulatory body
• Has NOT undergone clinical validation, regulatory review, or approval of any kind
• Is NOT intended to diagnose, treat, cure, or prevent any disease or medical condition
• Should NOT be used as a substitute for professional medical advice, diagnosis, or treatment
• Has NOT been validated in a prospective clinical setting
• Makes NO claims of clinical accuracy, safety, or efficacy

If you are experiencing a medical emergency or have concerns about a diabetic foot ulcer, contact your healthcare provider immediately.

Any use of this software for clinical decision-making is strictly prohibited and done entirely at the user's own risk. The authors, Northeastern University, and all affiliated parties disclaim all liability for any harm resulting from the use or misuse of this software.

For information on FDA-cleared wound measurement devices, visit FDA Medical Device Databases.

---

v1 → v2 Changelog

See CHANGELOG.md for the complete list of changes.

---

Citation

If you use this work for academic purposes, please cite:

@misc{bandari2026diafoot,
  title={DiaFoot.AI: A Multi-Task Pipeline for Diabetic Foot Ulcer Detection and Segmentation},
  author={Bandari, Ruthvik},
  year={2026},
  institution={Northeastern University},
  course={AAI6620 Computer Vision},
  note={Academic project — not for clinical use}
}

---

License

MIT License. See LICENSE for details.

This license grants permission for academic and research use. It does not grant permission for clinical or diagnostic use.

---

Built with care for educational impact. Data composition matters more than architecture.