Machine Learning-Based Biomarker for Enhanced Autism Diagnosis Using Fundus Images

Numerous children are suffering from neurodevelopmental disorders. According to the CDC, about 1 in 36 children were diagnosed as having ASD. The seriousness of this problem is also seen through the rapid increase in the prevalence of ASD with the combined prevalence per 1,000 children increasing from 6.7 in 2000 to 27.6 in 2020. The complete cure of neurodevelopmental disorders such as ASD is the key to early diagnosis of the disorder. One of the diagnostic methods using fundus images is an effective, non-invasive method; however, they need more research on how to capture key parts of the fundus images such as the optic nerve and the blood vessels. To address this problem, this study proposes a novel advanced image classification system with the use of an adversarial attack where the adversarial attack will be used to improve the function of the system and determine which parts of the fundus images are most crucial to getting an accurate result. To carry this out, the adversarial attack was applied to three distinct parts of the fundus image–the blood vessel isolated image, optic cup isolated image, and optic disc isolated image. Through extensive experiments, it is shown that the proposed method can be utilized as a biomarker for diagnosing ASD.