DiaFoot.AI — Diabetic Foot Ulcer Triage & Wound Segmentation

Cascaded multi-task model for diabetic foot images: it triages a foot photo into Healthy / Non-DFU / DFU, segments the wound when present, and measures wound area in mm². Both stages use a frozen DINOv2 ViT-B/14 backbone with a trainable head (classifier) / UPerNet decoder (segmenter).

Code: https://github.com/Ruthvik-Bandari/DiaFoot.AI

Honesty note. Every number below is from leakage-audited clean-split evaluation (results/*.json in the repo). They are lower than an earlier headline (Dice 85.89%) that was a DFU-only subgroup measured before a data-leakage fix. These are the honest numbers.

Files

File	Stage	Backbone
dinov2_classifier.pt	3-class triage (Healthy/Non-DFU/DFU)	DINOv2 ViT-B/14 + linear head
dinov2_segmenter.pt	Binary wound segmentation	DINOv2 ViT-B/14 + UPerNet decoder

Results (leakage-audited test set, n = 1,161)

Triage classification

Metric	Value
Accuracy	0.984
Macro F1	0.981
Macro AUROC	0.999
DFU sensitivity	0.966
Healthy specificity	0.995
ECE (after temperature scaling)	0.007
Defer @ 0.95 confidence	93.5% coverage, 99.7% accuracy on kept

Wound segmentation

Slice	Dice	IoU	HD95 (px)
DFU wounds only (n = 263)	0.891	0.829	11.3
Full mixed test set, mean	0.718	0.673	66.1
Full mixed test set, median	0.929	0.868	5.0
5-fold CV (DFU)	0.853 ± 0.009	0.785 ± 0.010	—

The mixed-set mean is far below the median because healthy/non-DFU images have empty masks (any false-positive pixel scores Dice ≈ 0). Judge wound quality from the DFU-only / median rows.

Intended use

Research and education on diabetic-foot-ulcer imaging. Not a medical device; no regulatory clearance. Do not use for diagnosis or treatment.

Limitations (read before use)

• The triage classifier does not generalize across image sources. On unseen datasets, external accuracy drops to ~21% and DFU sensitivity to 0%. Re-validate on any new source before use. The segmenter transfers well (external DFU Dice 0.893).
• Fairness is under-powered by the test split (effectively one ITA skin-tone group); the DFU-only fairness gap is 0.00, but broad skin-tone fairness is unproven.
• Clinical wound-area agreement was measured on only n = 3 (indicative, not validated).

How to use

import torch
from huggingface_hub import hf_hub_download

repo = "Ruthvik-Bandari/DiaFoot.AI"
clf = hf_hub_download(repo, "dinov2_classifier.pt")
seg = hf_hub_download(repo, "dinov2_segmenter.pt")
# Load with the model definitions in src/models/ from the GitHub repo:
#   DINOv2Classifier(backbone="dinov2_vitb14"), DINOv2Segmenter(backbone="dinov2_vitb14")
# Or run the end-to-end CLI:
#   python scripts/predict.py --image foot.jpg \
#     --classifier-checkpoint dinov2_classifier.pt \
#     --segmenter-checkpoint dinov2_segmenter.pt --device cpu

Input size 518×518 (DINOv2). See the GitHub repo for the full inference pipeline, FastAPI service, and ONNX export.

Training data

~8,105 images across three categories (2,119 DFU incl. AZH · 3,300 healthy · 2,686 non-DFU), aggregated from FUSeg, AZH, Kaggle, and Mendeley sources. Splits are 70/15/15, doubly stratified by ITA skin tone and class, and audited to zero train/test leakage (has_any_leakage: False). The raw images are not redistributed here (third-party licenses + patient privacy).

Citation

Bandari, R. DiaFoot.AI v2: Diabetic Foot Ulcer Detection, Segmentation & Wagner Staging. 2026. https://github.com/Ruthvik-Bandari/DiaFoot.AI

Built on DINOv2 (Meta AI, Apache-2.0). Fine-tuned weights released under MIT.