Deep Learning Research for Identification of Autism Spectrum Disorder


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
  • Ka Chun WONG (Supervisor)
  • Kay Chen Tan (External person) (External Co-Supervisor)
Award date2 Aug 2021


Facing the globally rising prevalence of autism spectrum disorder (ASD), the symptom-based diagnostic criteria struggle to address the urgent public health concern due to the global shortfalls in well-qualified professionals. With the recent advances in neuroimaging techniques and next-generation sequencing, the successful application of computer aided approaches has provided timely opportunities to relieve the tension in healthcare services. Through sharing and consolidating independent samples across different studies, leveraging the integrated data from multiple datasets becomes a silver-bullet solution for intelligent computational assistance.

However, most of the existing approaches are subject to relatively small datasets due to the sophisticated sampling criteria. As the sample size increases, the inflated accuracy could decline significantly. How to maintain reliable performance on large datasets is rarely discussed. Additionally, little efforts have been made to explore the remarkable and reproducible autistic biomarkers from data-driven outcomes for measuring autistic neurological functions.

This thesis focuses on developing effective models to address the above limitations from different perspectives. The proposed models aim to improve the evaluation and measurement for ASD. The major contributions of this thesis are summarized as follows. Firstly, the basic background knowledge on the related work including pathological investigations, use of digital technologies, neuroimaging-based computer aided diagnosis models, and computer aided autistic biomarker discovery is provided. The systematic review helps to understand the main tasks of this thesis consisting of neuroimaging-based computer aided diagnosis models and computer aided autistic biomarker discovery. 

Secondly, three effective computational models are proposed for the precise diagnosis of ASD and its related disorders. By exploiting reproducible connectivity patterns of ASD, a graph-based classification model using deep belief network with automatic hyperparameter-tuning technique is developed. Despite reliable performance achieved for ASD, this model cannot be applied to a small dataset or other ASD-related rare disorders by exploiting fruitful information from other related tasks and other modality data. To mitigate the substantial subject heterogeneity, a multi-task learning framework is presented for the joint diagnosis of ASD-related disorders by inducing coupling between tasks and modeling the optimal correlation structure for different tasks. The simulation results demonstrate that the closer correlation between disorders can enable more effective training of multi-task learning. Given the natural multimodal characteristics of medical images, the existing methods fail to take into account the fact that there is insufficient multimodal imaging data for training at present and the generic multimodal fusion strategies become less effective to characterize multi-modal variabilities across subjects. To solve the inherent bias in a single modality, a multimodal fusion framework is developed to exploit the spatiotemporal resolution complementarity of multimodal data for improving classification performance. 

Lastly, although neuroimaging techniques are effective to analyze neuronal activities of ASD, they cannot cover all usage scenarios in reality. The key roles of genetic and environmental factors have been realized to increase risk for developing ASD. For accelerating the clinical biomarker quantitative analysis of ASD, a multi-label classification model is proposed to prioritize autistic risk candidate genes and toxic chemicals based on multiple heterogeneous biological databases.

Based on several real-world datasets, the effectiveness and reliability of the proposed models are demonstrated through the comparison with state-of-the-art methods.  Some rules found in this thesis are consistent with the conclusions of other previous studies. A portion of the identified autistic biomarkers have been validated by the existing literatures.