Classification of Autism Spectrum Disorder (ASD) Using Deep Learning

Autism Spectrum Disorders (ASD) are neurological disorders characterized by deficits in cognitive skills, social and communicative behaviors. A common way of diagnosing ASD is by studying behavioral cues, such as hand/arm flapping, head banging, and spinning. This project aims to leverage computer vision and machine learning techniques to improve the efficiency and accuracy of diagnosing ASD through behavioral cues.

Dataset

The SSBD dataset can be downloaded from Roland Goecke's research page . The dataset contains about 25 videos of hand flapping, head banging, and spinning behaviors.

Data Preparation and Preprocessing

1. Data Download: The dataset is downloaded from Google Drive if not already present.
2. Data Extraction: The downloaded zip file is extracted.
3. Frame Extraction: Frames are extracted from videos at regular intervals, resized to 128x128 pixels, normalized to the range [0, 1], and stored.
4. Dataset Creation: A dataset of sequences (each containing 60 frames) is created along with their corresponding labels.
5. Train-Test Split: The dataset is split into training and testing sets.

Data Preparation and Preprocessing

ConvLSTM2D Model

• Initialization: Load and unzip the dataset, set constants (IMAGE_WIDTH, IMAGE_HEIGHT, SEQUENCE_LENGTH).
• Dataset Preparation: Extract 60 frames from each video, resize them.
• Model Structure:
  • 4 convolutional layers, each followed by MaxPooling3D.
  • Flatten layers and use a dense layer with softmax activation for classification.
  • Use categorical_crossentropy loss function and Adam optimizer.
  • Train for 10 epochs with a 75:25 train-validation split.

LRCN Model

• Initialization: Load and unzip the dataset, set constants (IMAGE_WIDTH, IMAGE_HEIGHT, SEQUENCE_LENGTH).
• Dataset Preparation: Extract 60 frames from each video, resize them.
• Model Structure:
  • 4 convolutional layers followed by MaxPooling2D.
  • Flatten layers, followed by an LSTM layer with 32 units to capture temporal dependencies.
  • Dense layer with softmax activation for multi-class prediction (3 classes).
  • Use categorical_crossentropy loss function and Adam optimizer.
  • Train for 10 epochs with a 75:25 train-validation split.

Results

The models were evaluated on the test set, and performance metrics were calculated to determine the accuracy of classifying the behaviors.