Abstract
Owing to the concern regarding lack of reliance on cognitive
profile-based diagnosis of dyslexia, we conducted an automated diagnosis exclusively based on the handwriting sample. Handwriting sample
of 54 children (36 males, 18 females) studying between classes 1st to
5
th identified with ‘strong evidence of risk’ on Dyslexia Screening TestJunior (DST-J) was collected. 14 Hindi words (i.e. 5 two letter words, 6
words with Matras (vowel signs), and 3 conjoined consonants (Sanyukt
Akshar)) were selected for this study on the basis of graded difficulty
level. These Hindi words share features such as matras, killer strokes
(halants), and Sanyukt Akshar. A total of 267 images, 164 from children
with dyslexia-dysgraphia and 103 from age-matched normal control, were
collected for this study. These images were resized to a fixed height of
113 pixels along with different width sizes depending on the image aspect
ratio. A random number of patches with size of 113 × 113 pixels were
generated from each image. A Convolutional Neural Network (CNN)
using Keras and Tensorflow was successful in automatically identifying
powerful features with average accuracy of (86.14 ± 1.02)%. The findings endorse deep learning approach in automated detection/diagnosis
of dyslexia-dysgraphia
profile-based diagnosis of dyslexia, we conducted an automated diagnosis exclusively based on the handwriting sample. Handwriting sample
of 54 children (36 males, 18 females) studying between classes 1st to
5
th identified with ‘strong evidence of risk’ on Dyslexia Screening TestJunior (DST-J) was collected. 14 Hindi words (i.e. 5 two letter words, 6
words with Matras (vowel signs), and 3 conjoined consonants (Sanyukt
Akshar)) were selected for this study on the basis of graded difficulty
level. These Hindi words share features such as matras, killer strokes
(halants), and Sanyukt Akshar. A total of 267 images, 164 from children
with dyslexia-dysgraphia and 103 from age-matched normal control, were
collected for this study. These images were resized to a fixed height of
113 pixels along with different width sizes depending on the image aspect
ratio. A random number of patches with size of 113 × 113 pixels were
generated from each image. A Convolutional Neural Network (CNN)
using Keras and Tensorflow was successful in automatically identifying
powerful features with average accuracy of (86.14 ± 1.02)%. The findings endorse deep learning approach in automated detection/diagnosis
of dyslexia-dysgraphia
Original language | English |
---|---|
Publication status | Accepted/In press - 12 Nov 2020 |
Event | The 25th IEEE International Conference on Pattern Recognition - Milan, Italy Duration: 10 Jan 2021 → 15 Jan 2021 Conference number: 25th https://www.micc.unifi.it/icpr2020/ |
Conference
Conference | The 25th IEEE International Conference on Pattern Recognition |
---|---|
Abbreviated title | ICPR2020 |
Country/Territory | Italy |
City | Milan |
Period | 10/01/21 → 15/01/21 |
Internet address |
Keywords
- Deep Learning
- Dyslexia
- D