Autism Face Classification with Deep Learning

This repository presents a complete, end-to-end deep learning pipeline for binary classification of facial images related to Autism Spectrum Disorder (ASD).
The project emphasizes clean engineering, reproducibility, data-leakage–safe evaluation, and explainability, making it suitable for both academic research and professional portfolios.

---

🔍 Problem Overview

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that can affect social interaction and communication.
Early detection plays a crucial role in intervention strategies.

This project investigates whether convolutional neural networks (CNNs) can learn discriminative facial features to distinguish between:

• ASD (Autistic)
• TD (Typically Developing)

from static facial images.

---

⚠️ This project is research-oriented and not intended for clinical diagnosis.

---

🧠 Methodology

Model Architectures

ImageNet-pretrained CNN backbones are employed:

• ResNet50
• DenseNet121
• MobileNetV3-Small

A model factory pattern enables seamless architecture switching.

Training Strategy

• Transfer learning
• Optional backbone freezing (feature extraction vs full fine-tuning)
• Physical K-Fold Cross-Validation (K = 5)
  • Each fold is stored in separate directories
  • Guarantees zero data leakage between train and validation sets

Data Augmentation

Implemented using Albumentations:

• Resize
• Horizontal flip
• Small rotations
• Brightness & contrast adjustment
• ImageNet normalization

---

📁 Dataset Organization

The dataset is not included due to privacy and licensing constraints.

Expected structure:

data/autism_unified_kfold/
  fold_1/
    train/
      autistic/
      typical/
    val/
      autistic/
      typical/
  fold_2/
  ...

Physical fold separation is intentionally used instead of logical splits.

---

📊 Quantitative Results

5-Fold Cross-Validation Summary(DenseNet121)

Metric	Mean ± Std (%)
Accuracy	86.48 ± 2.48
Precision	86.70 ± 2.29
Recall	86.48 ± 2.48
F1-score	86.45 ± 2.51

Metrics are macro-averaged and computed on validation sets only.

---

Confusion Matrix (Example Fold)

The confusion matrix highlights class-wise performance and common misclassification patterns.

---

🧪 Explainability — Grad-CAM

To improve interpretability, Grad-CAM is used to visualize spatial regions that most influence model predictions.

This confirms that the network focuses on meaningful facial regions, rather than background artifacts.

---

📈 Training & Evaluation Pipeline

Pipeline highlights:

• Centralized configuration system
• Deterministic training via fixed seeds
• Automatic checkpointing (best validation loss)
• JSON-based experiment logging
• Script-based execution (no notebook dependency)

---

🚀 How to Run

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Train (Physical K-Fold)

python scripts/train.py   --data_dir ./data/autism_unified_kfold   --model resnet50   --epochs 30   --batch 32

Outputs:

• saved_models/ → fold-wise checkpoints
• outputs/ → logs, summaries, metrics

---

3️⃣ Evaluate a Trained Model

python scripts/eval.py   --data ./data/autism_unified_kfold/fold_1/val   --model resnet50   --weights ./saved_models/resnet50_fold1.pth   --out ./outputs/eval_fold1

Exports:

• metrics.json
• confusion_matrix.png

---

📓 Notebooks

Purpose-specific notebooks only (no core logic):

• 01_colab_runner.ipynb — Minimal Colab execution
• 02_gradcam.ipynb — Grad-CAM visualization

---

🛠️ Project Structure

src/
  config.py        # Experiment configuration
  dataset.py       # Dataset & augmentation pipeline
  models.py        # CNN model factory
  train.py         # K-Fold training engine
  eval.py          # Evaluation & reporting
  utils.py         # Utilities (seed, metrics)

scripts/
  train.py         # Training entrypoint
  eval.py          # Evaluation entrypoint

---

⚠️ Reproducibility Notes

• All experiments are configuration-driven
• Random seeds fixed for deterministic behavior
• num_workers=0 ensures cross-platform stability

---

📜 License

This project is released under the MIT License.

---

👤 Author

Vefa
Computer Engineering Student
Focus areas: Deep Learning, Computer Vision, Explainable AI

📌 Disclaimer

This project is intended for research and educational purposes only
and is not a clinical diagnostic tool.