Through this research presentation, we propose an autism spectrum disorder (ASD) prediction model for children using facial morphology in images and a deep convolutional neural network (via transfer learning), which collaboratively quantify the binary classification problem of distinguishing subjects with ASD from healthy subjects based on unconstrained two-dimensional images.
The corresponding oral presentation was presented at the IEEE EMBS International Student Conference 2021 - Moratuwa on 6th February 2021.
Abstract Link: http://dx.doi.org/10.13140/RG.2.2.36128.30724
Synopsis Presentation On Gsm based flood notification system
Similar to Predicting Autism Spectrum Disorder in Children based on Facial Morphological Characteristics: A Deep Convolutional Neural Network Framework
Face Mask and Social Distance DetectionIRJET Journal
Similar to Predicting Autism Spectrum Disorder in Children based on Facial Morphological Characteristics: A Deep Convolutional Neural Network Framework (20)
2. AUTIDET
AUTIDET is a deep convolutional
neural network framework to
predict and bilaterally classify
Autism Spectrum Disorder in
children based on facial
morphological characteristics.
AUTIDET is integrated with
a front-end mobile application with
a user-ergonomic interface to
process user input facial images of
children.
AutiDET
F o r Z e r o U n s c r e e n e d
3. repetitive behaviors
narrow interests
deficits in social interaction and
communication ability
A neurodevelopmental disorder
characterized by,
The cranio-facial anomalies frequently
occur in Autism Spectrum Disorder
especially in children
Autism Spectrum Disorder
4. [Elsabbagh M. et al., 2012]
in Global Context
Children has Autism Spectrum Disorder (ASD)
One in 160
5. 1.07%
of children in Sri Lanka has ASD
Higher prevalence in South Asian Region
than in Global Context
[Hossain M.D. et al., 2017]
Inadequate solid established set-up for
screening ASD in children in Sri Lanka
6. Constraints in
current settings
Identification of ASD through facial
phenotypes in clinical environments:
Exorbitant in cost
Time-consuming
Requirement of clinical expertise of a
psychiatrist or a neurologist
These factors buffer for early diagnosis and
intervention of ASD
Thus, reduce the outcomes from
therapeutic applications
https://www.sciencedirect.com/science/article/pii/S1071909120300425
7. The Objective
A novel approach based on neural networks to efficiently search
the neural markers of anomalous cranio-facial processing for the
early detection and diagnosis of ASD
8. Deep Convolutional
Neural Network (DCNN)
for Facial Morphology
For quantification of the
binary classification of
distinguishing subjects with
ASD from healthy subjects
based on facial images
An integrated front-end
mobile application with a
user-friendly interface
For the integration of DCNN
model into real world
scenario by predicting ASD in
user input facial images of
children
Our Approach
9. AutiDET
F o r Z e r o U n s c r e e n e d
DCNN for
Facial Morphology
DCNN model: via DenseNet-201
With the deep of 201 layers of depth-
wise convolutions
DCNN model: via MobileNet V-1
With the 28 separate layers
Dataset from Kaggle
Balanced with train, validation and test
datasets
10. AutiDET
F o r Z e r o U n s c r e e n e d
Dataset from
Kaggle
An annotated dataset of the autistic and non-
autistic facial images of children
Re-organized into:
training (~70%)
test (~20%)
validation (~10%) sub-datasets
Total collection of 2936 facial images of
children
Unconstrained two-dimensional images
[Kaggle.com, “Detect autism from a facial image,” 2020]
11. AutiDET
F o r Z e r o U n s c r e e n e d
DCNN model:
via DenseNet-201
Integrated model network architecture of
DenseNet 201
With convolutional and dense layers
12. AutiDET
F o r Z e r o U n s c r e e n e d
DCNN model:
via DenseNet-201
Receiver Operating Characteristics curve of
the DCNN model based on DenseNet-201
AUC score - 93.8%
13. AutiDET
F o r Z e r o U n s c r e e n e d
DCNN model:
via MobileNet V-1
Integrated model network architecture of
MobileNet V-1
With convolutional layers and blocks
14. AutiDET
F o r Z e r o U n s c r e e n e d
DCNN model:
via MobileNet V-1
Receiver Operating Characteristics curve of
the DCNN model based on MobileNet V-1
AUC score - 95.4%
16. The Integrated
front-end Mobile
Application with a
user-friendly
interface
Key Features
User-friendly with
a simple interface
Accurate visual
representation
Fast & fully
automated algorithm
Supportive to
clinical applications
17. Back-end DCNN
model
Developed using Python
Jupyter Notebook upon
Tensorflow keras platform
Application
Programming
Interface and Firebase
Implemented using Python
and Flask for application
programming interface while
Firebase as the cloud storage
Mobile Application
Developed using Flutter
environment and android
8.1 emulator
Integration of Mobile Application
in front-end
19. AutiDET
F o r Z e r o U n s c r e e n e d
AUTIDET
AUTIDET is a deep convolutional neural
network framework, integrated with
a front-end user-ergonomic mobile application,
to predict and bilaterally classify Autism
Spectrum Disorder in children based on facial
morphological characteristics.
We thoroughly believe that AUTIDET will be a
game-changer in early detection and diagnosis
of ASD with a fast, accurate and fully-
automated approach
21. THANK YOU...
Children are obviously the future of humankind. Lets make them free from Autism from this approach of early
diagnosis and the technological precision medicine. So,
"No more Unscreened"