MrCzaro/Pressure_sore_cascade_classifier_YOLO

Pressure Sore Cascade Classifier — YOLO Weights

Eight YOLO classification weights for a 3-level hierarchical cascade pipeline that detects and stages pressure sores (pressure ulcers) from clinical photographs.

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

---

Medical Context

Pressure sores (pressure ulcers) are localised injuries to skin and underlying tissue caused by prolonged pressure. They are staged I–IV based on severity:

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

Early detection and accurate staging are critical for treatment planning and preventing progression.

---

Cascade Pipeline

Instead of a single hard 4-class classifier, the pipeline decomposes staging into three sequential binary decisions. Each model only needs to learn one boundary at a time, which reduces inter-class confusion — especially between Stage III and IV:

Image
  │
  ▼
[Level 1]  Pressure Sore vs No Pressure Sore
  │ NO  → "No pressure sore detected"
  │ YES ↓
[Level 2]  Early (Stage I/II) vs Advanced (Stage III/IV)
  ├─ EARLY    → [Level 3a]  Stage I vs Stage II
  └─ ADVANCED → [Level 3b]  Stage III vs Stage IV

Each level uses a 2-model ensemble (averaged softmax probabilities) from two architecturally distinct YOLO families for maximum diversity and robustness.

---

Files

File	Size	Level	Task	Accuracy	AUC-ROC
Level 1 Binary PS or not PS YOLO8x.pt	112 MB	L1	PS vs No-PS	1.0000	1.0000
Level 1 Binary PS or not PS YOLO26x.pt	57 MB	L1	PS vs No-PS	1.0000	0.9991
Level 2 Early vs Advanced YOLO8m.pt	32 MB	L2	Early vs Advanced	0.9615	0.9948
Level 2 Early vs Advanced YOLO26x.pt	57 MB	L2	Early vs Advanced	0.9615	0.9969
Level 3a Early YOLO8n.pt	3 MB	L3a	Stage I vs II	1.0000	0.9501
Level 3a Early YOLO26x.pt	57 MB	L3a	Stage I vs II	0.9286	0.9521
Level 3b Advanced YOLO8x.pt	112 MB	L3b	Stage III vs IV	0.9286	0.9126
Level 3b Advanced YOLO26x.pt	57 MB	L3b	Stage III vs IV	0.9286	0.8366

---

Performance

Per-Level Ensemble Results

Level 1 — Binary: PS vs No-PS

Model	Accuracy	Macro F1	AUC-ROC
YOLO8x	1.0000	1.0000	1.0000
YOLO26x	1.0000	1.0000	0.9991
Ensemble	1.0000	1.0000	—

Level 2 — Binary: Early (Stage I/II) vs Advanced (Stage III/IV)

Model	Accuracy	Macro F1	AUC-ROC
YOLO8m	0.9615	0.9615	0.9948
YOLO26x	0.9615	0.9615	0.9969
Ensemble	0.9615	0.9616	—

Level 3a — Binary: Stage I vs Stage II (Early branch)

Model	Accuracy	Macro F1	AUC-ROC
YOLO8n	1.0000	1.0000	0.9501
YOLO26x	0.9286	0.9282	0.9521
Ensemble	0.9643	0.9644	—

Level 3b — Binary: Stage III vs Stage IV (Advanced branch)

Model	Accuracy	Macro F1	AUC-ROC
YOLO8x	0.9286	0.9282	0.9126
YOLO26x	0.9286	0.9282	0.8366
Ensemble	0.9286	0.9289	—

End-to-End Path Analysis

Joint F1 = product of ensemble F1 scores along each path — a conservative lower bound on end-to-end performance since errors at each level compound:

Cascade Path	L1 F1	L2 F1	L3 F1	Joint F1
PS → Early → Stage I or II	0.9981	0.9616	0.9644	0.9256
PS → Advanced → Stage III or IV	0.9981	0.9616	0.9289	0.8916

Both paths exceed the flat 4-class Torchvision baseline (0.74 macro F1). The cascade approach is most impactful at Level 3b — isolating the Stage III/IV boundary from unrelated Stage I/II samples is the primary driver of the performance gain on the hardest classification pair.

---

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_cascade_classifier_YOLO

Option B — download individual files in Python

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="MrCzaro/Pressure_sore_cascade_classifier_YOLO",
    filename="Level 1 Binary PS or not PS YOLO8x.pt"
)

3. Wire weights into the application

Clone the project repo and update the path constants in ps_classifier_yolo_cascade.py:

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

# ps_classifier_yolo_cascade.py — update these constants

L1_MODEL_PATHS = [
    "path/to/Level 1 Binary PS or not PS YOLO8x.pt",
    "path/to/Level 1 Binary PS or not PS YOLO26x.pt",
]
L2_MODEL_PATHS = [
    "path/to/Level 2 Early vs Advanced YOLO8m.pt",
    "path/to/Level 2 Early vs Advanced YOLO26x.pt",
]
L3_EARLY_MODEL_PATHS = [
    "path/to/Level 3a Early YOLO8n.pt",
    "path/to/Level 3a Early YOLO26x.pt",
]
L3_ADVANCED_MODEL_PATHS = [
    "path/to/Level 3b Advanced YOLO8x.pt",
    "path/to/Level 3b Advanced YOLO26x.pt",
]

Then run the application and select YOLO Cascade from the backend dropdown:

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

4. Download all weights and run cascade in one script

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

REPO = "MrCzaro/Pressure_sore_cascade_classifier_YOLO"

# Download all 8 weights
paths = {
    "l1": [
        hf_hub_download(REPO, "Level 1 Binary PS or not PS YOLO8x.pt"),
        hf_hub_download(REPO, "Level 1 Binary PS or not PS YOLO26x.pt"),
    ],
    "l2": [
        hf_hub_download(REPO, "Level 2 Early vs Advanced YOLO8m.pt"),
        hf_hub_download(REPO, "Level 2 Early vs Advanced YOLO26x.pt"),
    ],
    "l3_early": [
        hf_hub_download(REPO, "Level 3a Early YOLO8n.pt"),
        hf_hub_download(REPO, "Level 3a Early YOLO26x.pt"),
    ],
    "l3_advanced": [
        hf_hub_download(REPO, "Level 3b Advanced YOLO8x.pt"),
        hf_hub_download(REPO, "Level 3b Advanced YOLO26x.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")

# Level 1
_, l1_label, l1_conf = ensemble_predict(paths["l1"], img)
if "not" in l1_label.lower():
    print(f"No pressure sore detected ({l1_conf:.2f})")
else:
    # Level 2
    l2_idx, l2_label, l2_conf = ensemble_predict(paths["l2"], img)
    is_early = "early" in l2_label.lower()

    # Level 3
    branch = "l3_early" if is_early else "l3_advanced"
    _, l3_label, l3_conf = ensemble_predict(paths[branch], img)

    print(f"Pressure sore detected")
    print(f"Severity group : {l2_label} ({l2_conf:.2f})")
    print(f"Stage          : {l3_label} ({l3_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