Early detection of Autism Spectrum Disorder (ASD) plays a critical role in improving developmental outcomes.
This project implements a deep learning–based computer vision pipeline to classify facial images as Autistic or Non-Autistic using Convolutional Neural Networks (CNNs) and transfer learning.
Designed with research-level discipline and industry-grade structure, this repository demonstrates a complete end-to-end ML workflow.
- Binary image classification: Autistic vs Non-Autistic
- Facial image analysis using CNNs
- Transfer learning with Xception and VGG16
- Clean dataset handling with train / validation / test split
- Evaluation using standard medical ML metrics
- ✅ Pre-trained CNN models (ImageNet weights)
- ✅ Modular and reproducible pipeline
- ✅ Early stopping to prevent overfitting
- ✅ Detailed performance evaluation
- Source: Kaggle
- Name: Autism Image Dataset
- Classes:
- Autistic
- Non_Autistic
The dataset is already split to avoid data leakage and ensure fair evaluation.
- Pre-trained on ImageNet
- Frozen convolutional base
- Custom fully connected classifier head
- Classic deep CNN architecture
- Used for comparative performance analysis
- Image Size:
224 × 224 - Batch Size:
32 - Optimizer:
RMSprop - Loss Function:
Categorical Crossentropy - Callbacks:
EarlyStopping (validation loss)
- Accuracy
- Precision
- Recall (Sensitivity)
- F1-Score
- Confusion Matrix
These metrics provide a balanced and reliable evaluation, especially important for healthcare-related ML tasks.
The Xception-based model achieves high classification accuracy and strong generalization on unseen test data, validating the effectiveness of transfer learning for autism detection from facial images.
Results may vary slightly depending on hardware and random initialization.