Transformer-Based Deep Learning for Autism Detection from Facial Images

Faruk Cengiz; Fesih Keskin

Makale detayları

Transformer-Based Deep Learning for Autism Detection from Facial Images

Yazarlar:

Faruk Cengiz | Fesih Keskin

Yayıncı:

Igdir University

Ana Başlıklar:

Bilgisayar Yazılımı

Dergi:

Iğdır Üniversitesi Fen Bilimleri Enstitüsü Dergisi Vol.15 Issue 3, pp.755-764

Belge Tipi:

Makale

Yayın Tarihi:

1/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001356703

pISSN:

2536-4618

Anahtar kelimeler:

Özet:

 This study presents a comparative analysis of four different Transformer-based deep learning architectures Vision Transformer ViT, Swin Transformer Swin-T, Data-efficient Image Transformer DeiT, and Convolutional Transformer CoaT for Autism Spectrum Disorder ASD detection using facial images. In recent years, Transformer architectures have increasingly replaced traditional convolutional neural network based approaches in ASD detection research. In this context, experimental results demonstrate that the Swin-T model achieved the highest classification performance with 87.76 accuracy and an AUC of 0.96. The CoaT model followed closely with 86.01 accuracy and an AUC of 0.94, while DeiT 84.27 accuracy and ViT 82.52 accuracy exhibited relatively lower performance. Confusion matrix and ROC curve analyses confirm that the Swin-T model significantly reduced both false positive and false negative rates. These findings highlight the effectiveness of Swin-T and CoaT models in visual data processing and suggest that, when supported by larger datasets, these architectures could provide valuable contributions to early ASD diagnosis in both clinical and research domains.

 This study presents a comparative analysis of four different Transformer-based deep learning architectures Vision Transformer ViT, Swin Transformer Swin-T, Data-efficient Image Transformer DeiT, and Convolutional Transformer CoaT for Autism Spectrum Disorder ASD detection using facial images. In recent years, Transformer architectures have increasingly replaced traditional convolutional neural network based approaches in ASD detection research. In this context, experimental results demonstrate that the Swin-T model achieved the highest classification performance with 87.76 accuracy and an AUC of 0.96. The CoaT model followed closely with 86.01 accuracy and an AUC of 0.94, while DeiT 84.27 accuracy and ViT 82.52 accuracy exhibited relatively lower performance. Confusion matrix and ROC curve analyses confirm that the Swin-T model significantly reduced both false positive and false negative rates. These findings highlight the effectiveness of Swin-T and CoaT models in visual data processing and suggest that, when supported by larger datasets, these architectures could provide valuable contributions to early ASD diagnosis in both clinical and research domains.

 This study presents a comparative analysis of four different Transformer-based deep learning architectures Vision Transformer ViT, Swin Transformer Swin-T, Data-efficient Image Transformer DeiT, and Convolutional Transformer CoaT for Autism Spectrum Disorder ASD detection using facial images. In recent years, Transformer architectures have increasingly replaced traditional convolutional neural network based approaches in ASD detection research. In this context, experimental results demonstrate that the Swin-T model achieved the highest classification performance with 87.76 accuracy and an AUC of 0.96. The CoaT model followed closely with 86.01 accuracy and an AUC of 0.94, while DeiT 84.27 accuracy and ViT 82.52 accuracy exhibited relatively lower performance. Confusion matrix and ROC curve analyses confirm that the Swin-T model significantly reduced both false positive and false negative rates. These findings highlight the effectiveness of Swin-T and CoaT models in visual data processing and suggest that, when supported by larger datasets, these architectures could provide valuable contributions to early ASD diagnosis in both clinical and research domains.

 Bu çalışma, yüz görüntülerinden Otizm Spektrum Bozukluğu OSB tespiti amacıyla dört farklı Transformer tabanlı derin öğrenme mimarisinin Vision Transformer ViT, Swin Transformer Swin-T, Data-efficient Image Transformer DeiT ve Convolutional Transformer CoaT karşılaştırmalı analizini sunmaktadır. Son yıllarda, OSB tespitine yönelik araştırmalarda geleneksel evrişimsel sinir ağları tabanlı yaklaşımların yerini giderek Transformer mimarileri almaya başlamıştır. Bu kapsamda gerçekleştirilen deneyler, Swin-T modelinin 87,76 doğruluk ve 0,96 AUC ile en yüksek sınıflandırma performansına ulaştığını göstermektedir. CoaT modeli 86,01 doğruluk ve 0,94 AUC ile ikinci sırada yer alırken, DeiT 84,27 doğruluk ve ViT 82,52 doğruluk nispeten daha düşük başarı sergilemiştir. Karışıklık matrisi ve ROC eğrileri analizleri, Swin-T modelinin yanlış pozitif ve yanlış negatif oranlarını önemli ölçüde azalttığını ortaya koymaktadır. Elde edilen bulgular, özellikle Swin-T ve CoaT modellerinin görsel veri işleme konusundaki etkinliğini vurgulamakta ve bu mimarilerin daha büyük veri kümeleri ile desteklendiğinde erken OSB tanısı sürecine klinik ve araştırma alanlarında değerli katkılar sağlayabileceğini öne sürmektedir.

Makale detayları

Transformer-Based Deep Learning for Autism Detection from Facial Images

Yazarlar:

Faruk Cengiz | Fesih Keskin

Yayıncı:

Igdir University

Ana Başlıklar:

Bilgisayar Yazılımı

Dergi:

Iğdır Üniversitesi Fen Bilimleri Enstitüsü Dergisi Vol.15 Issue 3, pp.755-764

Belge Tipi:

Makale

Yayın Tarihi:

1/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001356703

pISSN:

2536-4618

Autism spectrum disorder

Vision transformers

Swin transformer

Deit

CoatOtizm spektrum bozukluğu

Vision transformers

Swin transformer

Deit

Coat

Sayı

Makale detayları

Yazarlar: Faruk Cengiz | Fesih Keskin

Yayıncı: Igdir University

Ana Başlıklar: Bilgisayar Yazılımı

Dergi: Iğdır Üniversitesi Fen Bilimleri Enstitüsü Dergisi Vol.15 Issue 3, pp.755-764

Belge Tipi: Makale

Yayın Tarihi: 1/2025

Belge Dili: Türkçe

Document ID (Org No.): 2001356703

pISSN: 2536-4618

Sayı

Yazarlar:

Faruk Cengiz | Fesih Keskin

Yayıncı:

Igdir University

Ana Başlıklar:

Bilgisayar Yazılımı

Dergi:

Iğdır Üniversitesi Fen Bilimleri Enstitüsü Dergisi Vol.15 Issue 3, pp.755-764

Belge Tipi:

Makale

Yayın Tarihi:

1/2025

Belge Dili:

Türkçe

Document ID (Org No.):

2001356703

pISSN:

2536-4618