Improving fMRI-Based Autism Severity Identification via Brain Network Distance and Adaptive Label Distribution Learning

Machine learning methodologies have been profoundly researched in the realm of autism spectrum disorder (ASD) diagnosis. Nonetheless, owing to the ambiguity of ASD severity labels and individual differences in ASD severity, current fMRI-based methods for identifying ASD severity still do not achieve satisfactory performance. Besides, the potential association between brain functional networks(BFN) and ASD symptom severity remains under investigation. To address these problems, we propose a low&high-level BFN distance method and an adaptive multi-label distribution(HBFND-AMLD) technique for ASD severity identification. First, a low-level and high-level BFN distance(HBFND) is proposed to construct BFN that reflects differences in ASD severity. This method can measure the distance between the ASD and the health control(HC) on the low-order and high-order BFN respectively, which can distinguish the severity of ASD. After that, a multi-task network is proposed for ASD severity identification which considers the individual differences of ASD severity in communication and society, which considers the individual differences in language and social skills of ASD patients. Finally, a novel adaptive label distribution(ALD) technique is employed to train the ASD severity identification model, effectively preventing network overfitting by restricting label probability distribution. We evaluate the proposed framework on the public ABIDE I dataset. The promising results obtained by our framework outperform the state-of-the-art methods with an increase in identification performance, indicating that it has a potential clinical prospect for practical ASD severity diagnosis.