SSMDA: Semi-supervised multi-source domain adaptive autism prediction model using neuroimaging.

Autism spectrum disorder (ASD) is notoriously difficult to diagnose, despite having a high prevalence. Recently, existing studies have turned to neuroimaging to enhance the clinical relevance and effectiveness of diagnostic findings. While multi-site data offers the advantage of larger sample sizes, they also suffer from inter-site heterogeneity and a lack of labeled sample data. To address these issues, we introduced a semi-supervised multi-source domain adaptation-based model (SSMDA) that harnesses both structural and functional MRI data (sMRI and fMRI) to facilitate the early identification and treatment of ASD. The proposed approach incorporates established feature-level adaptation techniques, specifically Correlation Alignment (CORAL), Joint Distribution Adaptation (JDA), and Transfer Joint Matching (TJM). Operating in a semi-supervised manner, the model utilizes pseudo labeling in tandem with a supervised loss i.e. binary cross-entropy. Additionally, binary cross entropy serves as a regularization strategy during the training phase of feature adaptation. We validate our proposed technique first on existing DA methods such as JDA, TJM, and CORAL on the publically accessible dataset ABIDE and then by using binary cross-entropy in existing DA methods such as TJM using cross entropy namely TJMCE and JDA using cross entropy namely JDACE. The experimental results show that our proposed SSMDA_JDACE method can significantly outperform competitive methods for multi-center ASD diagnosis. Particularly in the context of sMRI, an accuracy of 80.25% and a sensitivity of 79.92% were achieved. Furthermore, in the domain of fMRI, the approach demonstrated a commendable accuracy of 85.89%, accompanied by a sensitivity of 79.83%. Experimental results demonstrate that our method yields superior performance over several state-of-the-art multisite ASD studies. • Presenting a novel semisupervised multisource DA model for improved robustness, generalization, and alignment. • Utilizes feature-level information from multiple labelled source domains via TJM, CORAL, and JDA. • Introduces binary cross entropy loss with existing DA methods for performance boost. • First time suggesting a DA based technique for multi-site structural MRI of ASD data. • Experiments were conducted to evaluate the usefulness of the proposed method. [ABSTRACT FROM AUTHOR]

Copyright of Biomedical Signal Processing & Control is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)