Early Autism Disorder Detection Through Visualizing Eye-Tracking Patterns Using Compact Convolutional Transformers

What is it about?

Autism spectrum disorder (ASD) encompasses a wide range of conditions, including impairments in social interaction, communication, and relationships, as well as unusual behavioral patterns. Multiple new computer-aided technologies have improved autism diagnosis. Fifty-nine youngsters in the school age range participated in the research, during which they were presented with a variety of age-appropriate photographs and video recordings focusing on social behavior. Calibration and data collection on eye movement were performed on 13 youngsters with autism. Only two of the six people who failed to finish the exercise experienced calibration problems, but four of them failed to even peek at the screen. The InceptionV3 model improved accuracy to 68.8 percent, ranking it above both the EfficientNetB7 and MobileNetV2 models. As a result, Xception, a variant of the Inception architecture that adds new features, while excluding depth-wise separable convolutions, has been put into practice. After that, the Visual Transformer Model performed admirably, reaching 91% accuracy. The parameter of the Compact Convolutional Transformer (CCT) is 0.4M, which is quite tiny in comparison to the parameters of other models. Adam optimizer was released with 32 batches and a 30% dropout rate. In all, 212 ASD samples were correctly identified as positive, whereas 7 were misclassified as non-ASD pictures. While the original Inception design had depth-separable convolutions, the Xception variant adds this feature. Xception outperforms any previously deployed model in terms of accuracy. With an accuracy of 87.22 percent, VGG16 is superior to Xception. The Overall accuracy was 96.43 percentage points, with a recall of 96.29 and an F1 score of 96.0 percent.

Featured Image

Why is it important?

ASD comprises several impairments, including social, communication, and relationship difficulties and abnormal behaviours. Innovative computer-aided autism diagnostic approaches exist. The research provided 59 schoolchildren with age-appropriate social conduct photographs and videos. Eye movement calibration and data collection were done on 13 autistic youngsters. Only two of the six who failed had calibration concerns, but four never glanced at the screen. The 68.8% accurate InceptionV3 model beat EfficientNetB7 and MobileNetV2. Thus, Xception, a variant of Inception that adds features but removes depth-wise separable convolutions, has been constructed. Visual Transformer Model worked effectively with 91% accuracy. The 0.4M Compact Convolutional Transformer (CCT) parameters are tiny compared to other types. Adam optimised 32 batches with 30% dropout. Seven non-ASD pictures were misclassified, but 212 were positive. Inception contains depth-separable convolutions; Xception adds them. The accuracy of Xception surpasses all previous models. VGG16 is more accurate than Xception at 87.22 percent. F1 score 96.0 percent, accuracy 96.43 percent, recall 96.29 percent.

Perspectives