CareGuide AI — PulmoScan Chest · SkinGuide Dermatology

Health education & triage for chest X-rays, CT slices, and skin photos

NIH 112K · CheXpert 224K · HAM10000 10K · Emergency gate · Grad-CAM · Docker-ready

---

Overview

CareGuide AI (evolved from PulmoScan AI) is a production-style health education platform. It helps users understand possible findings in chest imaging and skin photos when a clinician is not immediately available — not a replacement for doctors and not for emergencies.

⚠️ Disclaimer: Phase 1 wellness/education tool. Not FDA-cleared. Call 911 for emergencies.

Repository: github.com/yashwanth123/PulmoScan-AI

git clone https://github.com/yashwanth123/PulmoScan-AI.git
cd PulmoScan-AI
git checkout cursor/pulmoscan-ai-platform-1c3d

---

What's New (v3.0)

#	Feature	Description
1	CareGuide rebrand	PulmoScan (chest) + SkinGuide (skin) modules
2	Emergency gate	Blocks triage when emergency symptoms reported
3	HAM10000 pipeline	Auto-download ~10K skin images + train_skin.py
4	NIH ChestX-ray14	Auto-download 112K label metadata + multi-label trainer
5	CheXpert support	Manual download validator for 224K chest X-rays
6	Educational triage UI	Patient-friendly guidance, not diagnosis language
7	Dataset registry API	GET /api/datasets lists all public data sources
8	Local test script	scripts/run_local_test.sh for Mac/CI smoke tests
9	Legacy COVID binary	Still supported via ml_training/train.py
10	Validation framework	ml_training/validate_datasets.py reports readiness

---

Massive Public Datasets

Dataset	Scale	Auto-download	Training script
NIH ChestX-ray14	112,120 X-rays	Metadata yes · images 30GB+	ml_training/train_chest_multilabel.py
CheXpert	224,316 X-rays	Manual (Stanford RUA)	Same + place under data/chexpert/
HAM10000	10,015 skin images	Yes (~5 GB)	ml_training/train_skin.py
education454 COVID	~2,300 X-rays	Yes (~1.3 GB)	ml_training/train.py (legacy)

# List all datasets
python3 scripts/download_datasets.py --list

# Download skin + NIH metadata + legacy COVID data
python3 scripts/download_datasets.py --dataset all

# HAM10000 only (~5 GB)
python3 scripts/download_datasets.py --dataset ham10000

# Validate what's ready for training
python3 ml_training/validate_datasets.py

CheXpert (manual — 439 GB)

1. Register at Stanford CheXpert
2. Extract to data/chexpert/ with train.csv, valid.csv, and train/ images
3. Validate: python3 scripts/download_datasets.py --dataset chexpert --validate-only

---

Project Structure

PulmoScan-AI/
├── backend/app/
│   ├── modules/              # Emergency gate + educational triage
│   ├── api/routes/           # predict, emergency, health, samples
│   └── ml/                   # EfficientNet, Grad-CAM, inference
├── ml_training/
│   ├── datasets/             # HAM10000, NIH, CheXpert downloaders
│   ├── train.py              # Legacy COVID binary
│   ├── train_skin.py         # SkinGuide on HAM10000
│   ├── train_chest_multilabel.py
│   └── validate_datasets.py
├── scripts/
│   ├── download_datasets.py  # Master dataset CLI
│   └── run_local_test.sh     # Local smoke tests
├── frontend/                 # CareGuide UI + emergency modal
└── tests/

---

Quick Start (Mac)

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt -r requirements-dev.txt

# Optional: download datasets
python3 scripts/download_datasets.py --dataset ham10000

# Train (pick one)
python3 ml_training/train.py --epochs 25              # legacy COVID binary
python3 ml_training/train_skin.py --quick             # skin smoke test
python3 ml_training/train_chest_multilabel.py --quick # chest multi-label smoke

python3 scripts/setup_samples.py
python3 run.py
# → http://localhost:8000

---

Testing

make test-fast          # Skip slow TensorFlow inference
make test               # Full suite
make test-local         # Emergency API + datasets + pytest
bash scripts/run_local_test.sh

API endpoints (new)

Method	Endpoint	Description
GET	/api/emergency/symptoms	Emergency checklist
POST	/api/emergency/screen	Pre-upload emergency gate
GET	/api/modules	Chest + skin modules
GET	/api/datasets	Public dataset registry

---

Regulatory Roadmap (US)

Phase	Scope
1 (now)	Wellness/education disclaimers, emergency gate, no diagnosis claims
2	Clinician-in-the-loop pilots, audit logs, label validation
3+	FDA 510(k) as CADe/triage aid — not autonomous diagnosis

---

License

MIT — research and educational use only. Not for clinical diagnosis or emergency care.