MrCzaro/Pressure_sore_classifier_YOLO

Pressure Sore Classifier — YOLO 2-Stage Weights

Six YOLO classification weights for a 2-stage cascade pipeline that detects and stages pressure sores from clinical photographs. Stage 1 determines whether a pressure sore is present; Stage 2 classifies its severity (Stage I–IV).

Full project code, training notebooks, and web application:
👉 github.com/MrCzaro/PS_Classifier

For the 3-level hierarchical cascade variant (8 models, binary decisions at every level):
👉 MrCzaro/Pressure_sore_cascade_classifier_YOLO

---

Medical Context

Pressure sores are staged I–IV based on tissue damage depth:

Stage	Definition	Key Visual Feature
I	Non-blanchable erythema, intact skin	Redness, no open wound
II	Partial-thickness skin loss, exposed dermis	Shallow open ulcer or blister
III	Full-thickness skin loss, fat visible	Deep crater, no exposed bone/tendon
IV	Full-thickness tissue loss, muscle/bone exposed	Exposed deep structures, eschar/slough

---

Pipeline

Image
  │
  ▼
[Stage 1]  Binary Detection — PS vs No-PS
           3-model ensemble: YOLO11s + YOLOv8x + YOLO26x
  │ NO  → "No pressure sore detected"
  │ YES ↓
[Stage 2]  Multiclass Staging — Stage I / II / III / IV
           3-model ensemble: YOLOv8n + YOLO26m + YOLOv8x

Both stages use averaged softmax probabilities across three architecturally distinct YOLO families (v8, YOLO11, YOLO26) for robustness.

---

Files

File	Stage	Task	Accuracy	AUC-ROC
Binary_YOLOv8x.pt	S1	PS vs No-PS	1.0000	0.9998
Binary_YOLOv11s.pt	S1	PS vs No-PS	1.0000	1.0000
Binary_YOLOv26x.pt	S1	PS vs No-PS	1.0000	1.0000
Multiclass_YOLOv8n.pt	S2	Stage I–IV	0.8571	0.9486 (macro OvR)
Multiclass_YOLOv8x.pt	S2	Stage I–IV	0.8571	0.9306 (macro OvR)
Multiclass_YOLOv26m.pt	S2	Stage I–IV	0.8571	0.9247 (macro OvR)

---

Performance

Stage 1 — Binary Detection (3-Model Ensemble)

Model	Accuracy	Macro F1	AUC-ROC
YOLO11s	1.0000	1.0000	0.9998
YOLOv8x	1.0000	1.0000	1.0000
YOLO26x	1.0000	1.0000	1.0000
Ensemble	1.0000	1.0000	—

Stage 2 — Multiclass Staging (3-Model Ensemble)

Model	Accuracy	Macro F1	Macro AUC-ROC (OvR)
YOLOv8n	0.8571	0.8500	0.9486
YOLO26m	0.8571	0.8542	0.9306
YOLOv8x	0.8571	0.8501	0.9247
Ensemble	0.8571	0.8514	—

Per-Class AUC-ROC (Stage 2)

Stage	YOLOv8n	YOLO26m	YOLOv8x
Stage I	0.9913	0.9712	0.9934
Stage II	0.9455	0.9261	0.9268
Stage III	0.8852	0.8866	0.8450
Stage IV	0.9723	0.9386	0.9337

The 3-model staging ensemble achieves 0.8514 Macro F1 — a significant improvement over the flat Torchvision staging baseline (0.74 Macro F1) used in Backend 1 of the same project.

Model Selection Rationale

Stage 1 — all three models achieve perfect or near-perfect accuracy and AUC; the three were chosen from different YOLO families (v8, YOLO11, YOLO26) to maximise ensemble diversity.

Stage 2 — selected by Macro AUC-ROC as primary criterion (threshold-independent, treats all 4 classes equally) with Macro F1 as tiebreaker:

Pick	Model	Reason
1	YOLOv8n	Best AUC+F1 balance, strongest Stage IV AUC (0.9723)
2	YOLO26m	Best Macro F1 overall, best Stage III AUC among YOLO26 models, architecture diversity
3	YOLOv8x	Third AUC, strong Stage I AUC (0.9934), large v8 model

---

Installation & Usage

1. Install dependencies

pip install ultralytics huggingface_hub

2. Download weights

Option A — clone the full repo

git lfs install
git clone https://huggingface.co/MrCzaro/Pressure_sore_classifier_YOLO

Option B — download individual files

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="MrCzaro/Pressure_sore_classifier_YOLO",
    filename="Binary_YOLOv8x.pt"
)

3. Wire weights into the application

git clone https://github.com/MrCzaro/PS_Classifier
cd PS_Classifier
pip install -r requirements.txt

Update path constants in ps_classifier_yolo.py:

BINARY_MODEL_PATHS = [
    "models/yolo/Binary_YOLOv8x.pt",
    "models/yolo/Binary_YOLOv11s.pt",
    "models/yolo/Binary_YOLOv26x.pt"
]
STAGE_MODEL_PATHS = [
    "models/yolo/Multiclass_YOLOv8n.pt",
    "models/yolo/Multiclass_YOLOv8x.pt",
    "models/yolo/Multiclass_YOLOv26m.pt"
]

Then run and select YOLO from the backend dropdown:

python main.py
# Open http://localhost:5001

4. Standalone inference script

from huggingface_hub import hf_hub_download
from ultralytics import YOLO
import numpy as np
from PIL import Image

REPO = "MrCzaro/Pressure_sore_classifier_YOLO"

BINARY_MODEL_PATHS = [
    "models/yolo/Binary_YOLOv8x.pt",
    "models/yolo/Binary_YOLOv11s.pt",
    "models/yolo/Binary_YOLOv26x.pt"
]
STAGE_MODEL_PATHS = [
    "models/yolo/Multiclass_YOLOv8n.pt",
    "models/yolo/Multiclass_YOLOv8x.pt",
    "models/yolo/Multiclass_YOLOv26m.pt"
]

def ensemble_predict(model_paths, img):
    all_probs, names = [], None
    for p in model_paths:
        model  = YOLO(p)
        result = model(img, verbose=False)[0]
        all_probs.append(result.probs.data.cpu().numpy())
        if names is None:
            names = result.names
    avg = np.mean(all_probs, axis=0)
    idx = int(np.argmax(avg))
    return idx, names[idx], float(avg[idx])

img = Image.open("your_image.jpg").convert("RGB")

# Stage 1 — binary detection
_, label, conf = ensemble_predict(BINARY_MODEL_PATHS, img)

if "not" in label.lower():
    print(f"No pressure sore detected ({conf:.2f})")
else:
    # Stage 2 — multiclass staging
    _, stage, stage_conf = ensemble_predict(STAGE_MODEL_PATHS, img)
    print(f"Pressure sore detected ({conf:.2f})")
    print(f"Stage: {stage} ({stage_conf:.2f})")

---

⚠️ Medical Disclaimer

This model is provided for research and educational purposes only.

• ❌ NOT a medical device
• ❌ NOT certified for clinical diagnosis
• ❌ NOT a substitute for professional medical judgment
• ❌ NOT validated in clinical trials

Always consult licensed healthcare professionals for medical diagnosis and treatment.

---

Contact

Author: MrCzaro
GitHub: @MrCzaro
Project: PS_Classifier
Email: cezary.tubacki@gmail.com
License: MIT