---

title: Deepsyndrome Api emoji: 🧬 colorFrom: indigo colorTo: cyan sdk: gradio app_file: app_hf.py pinned: false

DeepSyndrome — AI Diagnostic System

Ensemble of ResNet-50 + EfficientNet-B3 + ViT-S/16 for Down Syndrome detection from facial imagery.

---

Project Structure

deep-syndrome/
├── app.py                    # Entry point — starts the server
├── main.py                   # FastAPI app + model loading + API routes
│
├── train.py                  # Train ALL 3 models in one go
├── train_resnet50.py         # Train ResNet-50 only
├── train_efficientnet.py     # Train EfficientNet-B3 only
├── train_vit.py              # Train ViT-S/16 only
│
├── training/                 # Shared training utilities
│   ├── dataset.py            # Dataset class, transforms, dataloaders
│   ├── engine.py             # Train loop, early stopping, evaluation
│   └── plots.py              # Training curves, ROC, Grad-CAM
│
├── data/
│   └── Down Syndrome Dataset/
│       ├── downSyndrome/     # Down Syndrome images
│       └── healthy/          # Control images
│
├── static/
│   └── index.html            # Web UI (served at http://localhost:7860)
│
├── requirements.txt          # Server dependencies
├── requirements_train.txt    # Training dependencies
├── ResNet50_best.pth         # Trained ResNet-50 weights
├── EfficientNet_B3_best.pth  # Trained EfficientNet-B3 weights
├── ViT_S16_best.pth          # Trained ViT-S/16 weights
└── results/                  # Auto-created after training
    ├── resnet50/             # ResNet-50 curves, confusion matrix, Grad-CAM
    ├── efficientnet/         # EfficientNet-B3 plots
    └── vit/                  # ViT-S/16 plots

---

Quick Start — Run the Server

Pre-trained .pth files are already included. Just install and run.

Step 1 — Install server dependencies

pip install -r requirements.txt

Step 2 — Start the server

python app.py

Step 3 — Open the web UI

http://localhost:7860

Upload any facial image and get predictions from all 3 models instantly.

---

Train Your Own Models

Use this when you want to retrain on a new or updated dataset. After training finishes, the .pth files in the project root are automatically replaced.

Step 1 — Install training dependencies

pip install -r requirements_train.txt

Step 2 — Prepare your dataset

Your dataset folder should have subfolders named with these keywords:

Class	Accepted folder name keywords
Down Syndrome	down, syndrome, ds, positive, affected
Control (no DS)	control, normal, healthy, negative, non

Example valid structures:

my_dataset/
├── down_syndrome/     ← images here
└── control/           ← images here

my_dataset/
├── DS_images/
├── DS_more/
└── normal_faces/

The script auto-explores all subfolders recursively and splits 70% train / 15% val / 15% test.

Step 3 — Run training

Dataset is already included in data/Down Syndrome Dataset/.

Train all 3 models at once (recommended)

python train.py --data-dir "data/Down Syndrome Dataset"

Or train each model separately

python train_resnet50.py    --data-dir "data/Down Syndrome Dataset"
python train_efficientnet.py --data-dir "data/Down Syndrome Dataset"
python train_vit.py         --data-dir "data/Down Syndrome Dataset"

Folder picker dialog (no args needed)

python train.py

A window opens — select the data/Down Syndrome Dataset folder.

Step 4 — Per-model improvements

Model	Key improvement	Default LR	Default Epochs
ResNet-50	Label smoothing (0.1) + GridDistortion aug	1e-4	20
EfficientNet-B3	CLAHE + GaussianBlur augmentation	1e-4	20
ViT-S/16	Linear warmup (5ep) + CosineAnnealingLR	5e-5	25

All models share: early stopping (patience 7-8), AMP on GPU, CosineAnnealingLR.

Step 5 — Custom options (same flags for all scripts)

Flag	Default	Description
--data-dir	(dialog)	Path to dataset root folder
--output-dir	project root	Where to save .pth files
--epochs	model-specific	Training epochs
--batch-size	32	Batch size
--lr	model-specific	Learning rate
--no-gradcam	off	Skip Grad-CAM visualization
--workers	0	DataLoader workers (keep 0 on Windows)

Examples:

# Quick test
python train_resnet50.py --data-dir "data/Down Syndrome Dataset" --epochs 5

# ViT with custom LR
python train_vit.py --data-dir "data/Down Syndrome Dataset" --epochs 30 --lr 3e-5

Step 5 — Restart the server to load new weights

python app.py

Training outputs saved to results/:

• training_curves.png — loss / accuracy / AUC per epoch
• roc_curves.png — ROC curve comparison
• *_confusion.png — confusion matrix per model
• *_gradcam.png — Grad-CAM heatmaps (best model)
• metrics.csv — accuracy, F1, AUC-ROC table

---

API Endpoints

Method	Endpoint	Description
GET	/	Web UI
GET	/health	Model load status
GET	/api	API info JSON
POST	/predict	Run prediction — send multipart/form-data with file field
GET	/docs	Swagger UI (interactive API docs)

Example curl request:

curl -X POST http://localhost:7860/predict \
  -F "[email protected]"

Example response:

{
  "prediction": "Down Syndrome",
  "confidence": 87.43,
  "probabilities": {
    "control": 12.57,
    "down_syndrome": 87.43
  },
  "models_used": ["ResNet-50", "EfficientNet-B3", "ViT-S/16"],
  "individual": {
    "ResNet-50":       { "prediction": "Down Syndrome", "confidence": 91.2 },
    "EfficientNet-B3": { "prediction": "Down Syndrome", "confidence": 85.6 },
    "ViT-S/16":        { "prediction": "Control",       "confidence": 78.1 }
  },
  "demo_mode": false
}

---

Notes

• Images are processed in-memory and never saved to disk.
• If no .pth files are found, the server runs in Demo Mode (random predictions).
• On Windows, keep --workers 0 during training to avoid multiprocessing issues.
• CPU inference is supported. GPU (CUDA) is auto-detected and used if available.