We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
A hybrid framework for detection of autism using ConvNeXt-T and embedding clusters.
- Authors
Kanwal, Ayesha; Javed, Kashif; Ali, Sara; Rubab, Saddaf; Khan, Muhammad Attique; Alasiry, Areej; Marzougui, Mehrez; Shabaz, Mohammad
- Abstract
Autism spectrum disorder (ASD) is a neurodevelopment disorders, which is very difficult to diagnose at early stages of life, and is a complex spectrum disorder, involving many persistent challenges with lack of social interactions, eye contact and facial expressions. Facial features of children having ASD are different from typically developing (TD) children, which is a major clue for early detection of autism using demographics of facial features. In this paper, a novel framework for early detection of autism has been developed, using a semi-supervised approach that consists ConvNeXt-T model and clustering of image embedding vectors. State-of-the-art ConvNeXt-T model is trained for autism dataset, and after that, image embedding clusters are created based on cosine similarity threshold. By doing this, model's classification layer and embedding generation layers are trained for the subject data. The proposed framework provides embeddings and autism classifier from the trained model, where cluster embeddings can be used in semi-supervised classification. Hence, after training process, inference architecture has two modes: (i) classification using ConvNeXt-T and (ii) classification using embedding clusters. The classification accuracy of framework using ConvNeXt-T is 89.5%, where clustering accuracy using cosine similarity of image embedding is 93%. Most of the other techniques in this area focus only on classification using deep neural networks. The proposed framework is also providing classification based on embedding clusters and can work on unlabelled dataset. Moreover, the proposed framework is based on server–client architecture, which allows easy usability and integration without multi acceleration devices and extra resource consumption. A single image of child is required to detect autism using the proposed framework in real-life scenario.
- Subjects
ARTIFICIAL neural networks; CHILDREN with autism spectrum disorders; COSINE function; AUTISM spectrum disorders; AUTISM; EYE contact; FACIAL expression
- Publication
Journal of Supercomputing, 2024, Vol 80, Issue 6, p8156
- ISSN
0920-8542
- Publication type
Article
- DOI
10.1007/s11227-023-05761-8