AI systems paired with brain imaging may support earlier autism diagnosis, according to a peer-reviewed review, but the authors say the field still needs standardized protocols, external validation, explainable AI and clinically translatable frameworks.
The paper is therefore a report on the state of the research, not a model-specific clinical validation result. It is identified as a review article and uses a critical analysis of literature that includes experimental and review studies.
A map of the field
The review focuses on recent autism spectrum disorder research, especially neuroimaging and AI- and machine-learning-based diagnostic approaches. It also covers prevalence trends, diagnostic methods and therapeutic interventions, giving the article a broader scope than a study of one algorithm or one scan.
Its stated outputs are a synthesis of neuroimaging and AI-based diagnostic methods, an analysis of available datasets and a critical evaluation of methodological challenges and future directions. The review also summarizes publicly available ASD datasets as part of its evidence base.
That combination lets the authors examine not only what kinds of tools are being developed, but also the data and research practices those tools depend on. Their conclusion is cautious: AI-driven systems have potential to support early diagnosis, but that potential still has to be tested and translated into clinical frameworks.
Why the findings may not travel
The review identifies small sample sizes, limited population diversity, heterogeneous imaging protocols, restricted generalizability and reliance on single-modal datasets as key challenges. In ordinary terms, some studies draw on relatively narrow groups, use different imaging procedures or rely on one kind of data, making results harder to compare or carry into other settings.
Taken together, those problems make it difficult to know how consistently a result would hold when the population, research site or imaging setup changes. The supplied analysis does not quantify how frequently these challenges occur or how much they affect performance, so the review offers a field-level warning rather than a measured comparison of diagnostic accuracy.
The supplied analysis reports no model-specific performance estimates or clinical validation results. It also does not demonstrate generalizability across diverse populations, imaging protocols or multicenter settings.
The evidence still needed
The authors call for standardized protocols, external validation, explainable AI and clinically translatable frameworks. Explainable AI means making the basis for a system's output easier to inspect, while external validation tests whether an approach holds beyond the data used in its initial development or assessment.
Future work should use more diverse datasets and multicenter clinical validation. Testing approaches across different populations and sites would directly address the generalizability problem highlighted by the review.
The authors also prioritize adaptive learning methods to improve the reliability and applicability of ASD diagnostic systems. Several questions remain open: how well will these approaches validate across diverse and multicenter populations, which imaging and evaluation protocols will best support clinical translation, and whether explainable and adaptive methods can make diagnostic systems more dependable.
Publication record
The article was accepted on 09 August 2026 and published as the version of record on 21 August 2026. The authors declare no conflicts of interest.
Paper data and sources
Original title: AI and neuroimaging in autism spectrum disorder: advances in diagnosis, methodological challenges and future directions.
Authors: Kuljeet Singh, Khushi Mogha, Sidi Mohamed Sid'El Moctar
Journal/Repository: Neurological sciences : official journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology
Status: Peer-reviewed
First online: 2026-08-21
DOI: 10.1007/s10072-026-09334-4
Original paper · Full text