<p>Autism Spectrum Disorder or ASD is a developmental disorder that mainly affects the way an individual communicates and interacts with other people, besides exhibiting self-absorbing behaviours & showcasing intellect disability. ASD is said to have become very rampant, thus early and accurate diagnosis is essential in order to ensure that the appropriate help is given to the affected people. In the past, methods of ASD diagnosis have involved interviews and observations, which can be highly subjective, as well as requiring a large amount of time and effort to administer and may vary from assessor to assessor. These limitations indicate the importance of more accurate, less subjective, and standardized diagnostic tools. Recent enhancements of AI, especially, machine learning (ML) and deep learning, make Convolutional Neural Networks (CNNs) as a powerful tool to detect and classify ASD effectively. CNNs, a subclass of deep learning, are most effective in image and pattern analyzing which makes them useful in processing medical images like MRI or fMRI data or behaviors like gaze and facial movements. Since CNN is capable of automatically learning the input data’s hierarchical features, the accuracy of distinguishing ASD individuals from normal individuals is very high. Therefore, the objective of this paper is to present a systematic review of CNN-based approaches to the diagnosis of autism. In this section, we explain the types of CNN architecture adopted in the aforementioned studies, the obtained results, and limitations encountered in diagnosing ASD using CNNs. Some of the challenges include aspects such as large annotated datasets, interpretability of machine learning models and the ability to generalize to other population. To tackle the problem of opacity inherent in deep learning models, we highlight the emerging field of explanation and interpretable AI (XAI). XAI is particularly useful for identifying autism using the CNN since it helps in explaining model decisions, which are important in medical settings. Besides the previous approaches, this paper discusses the directions for the future research in the field of CNN-based autism detection, namely the utilization of the multimodal information and the improvement of the interpretability and robustness of the algorithm. We also draw attention to the possible implications of CNNs for clinical applications, but focusing on the significance of the early diagnosis of ASD since CNNs could provide an automated and quick means of diagnosing this disorder, which in turn would help to avoid delays in patient treatment.</p>

A comprehensive review of autism detection using deep learning

期刊论文
发表日期：

2026年4月13日

查看全文

M. Spoorthi,
Bipin Kumar Rai

摘要

Autism Spectrum Disorder or ASD is a developmental disorder that mainly affects the way an individual communicates and interacts with other people, besides exhibiting self-absorbing behaviours & showcasing intellect disability. ASD is said to have become very rampant, thus early and accurate diagnosis is essential in order to ensure that the appropriate help is given to the affected people. In the past, methods of ASD diagnosis have involved interviews and observations, which can be highly subjective, as well as requiring a large amount of time and effort to administer and may vary from assessor to assessor. These limitations indicate the importance of more accurate, less subjective, and standardized diagnostic tools. Recent enhancements of AI, especially, machine learning (ML) and deep learning, make Convolutional Neural Networks (CNNs) as a powerful tool to detect and classify ASD effectively. CNNs, a subclass of deep learning, are most effective in image and pattern analyzing which makes them useful in processing medical images like MRI or fMRI data or behaviors like gaze and facial movements. Since CNN is capable of automatically learning the input data’s hierarchical features, the accuracy of distinguishing ASD individuals from normal individuals is very high. Therefore, the objective of this paper is to present a systematic review of CNN-based approaches to the diagnosis of autism. In this section, we explain the types of CNN architecture adopted in the aforementioned studies, the obtained results, and limitations encountered in diagnosing ASD using CNNs. Some of the challenges include aspects such as large annotated datasets, interpretability of machine learning models and the ability to generalize to other population. To tackle the problem of opacity inherent in deep learning models, we highlight the emerging field of explanation and interpretable AI (XAI). XAI is particularly useful for identifying autism using the CNN since it helps in explaining model decisions, which are important in medical settings. Besides the previous approaches, this paper discusses the directions for the future research in the field of CNN-based autism detection, namely the utilization of the multimodal information and the improvement of the interpretability and robustness of the algorithm. We also draw attention to the possible implications of CNNs for clinical applications, but focusing on the significance of the early diagnosis of ASD since CNNs could provide an automated and quick means of diagnosing this disorder, which in turn would help to avoid delays in patient treatment.