City Research Online

Abstract

Attention Deficit Hyperactive Disorder (ADHD) is well-known common causation of childhood behavioural disorders. It is estimated that around 5-10% of children globally are affected with this disorder. ADHD is attributed to problematic behaviours that include inattention and impulsivity. Children find it extremely difficult to focus, be attentive and to organise themselves. It contributes to a lifetime of impairment, poor quality of life and long-term burden on affected families. Since there is no single cause found in the prevalence or absence of ADHD. The usual method of diagnosis is merely dependent on behavioural analysis which are all subjective. Clinicians usually take months to diagnose the condition. To date, there are no biological markers that exist for ADHD. To measure neurobiological data objectively, an assessment of the brain behaviour relationship is essential to transform the method of diagnosis. Automatic diagnosis is a profound way for an effective cure.

In this dissertation, we aim to solve the problem of automatic diagnosis of ADHD using machine learning methods based on functional MRI (fMRI) data. The proposed methods begin with classical machine learning and move to deep learning as a way to improve the classification performance. Interpretability of results is an important aspect, so functional connectivity is a central theme in the work and the proposed methods utilise functional connectivity in increasingly more complex ways.

In the first method, we have evaluated a clustering based novel method to calculate functional connectivity. After calculating functional connectivity, we employ Elastic Net feature selection to select the discriminant features and integrate non-imaging data. Finally, a Support Vector Machine (SVM) classifier is trained to classify ADHD.

The second method presents a deep learning based novel method, called FCNet, that calculates functional connectivity from fMRI time-series signals. The FCNet consists of two networks, i) a convolutional neural network in a Siamese architecture that extracts abstract features from a pair of time-series signals and, ii) a similarity measure network that computes the strength of similarity between the extracted features which serves as functional connectivity. Similar to the previous method, an Elastic Net and SVM is applied to classify ADHD.

In the third method, we have proposed an end-to-end trainable model to classify ADHD from preprocessed fMRI time-series data. The model takes fMRI time-series signals as input and outputs the predicted labels, and is trained end to-end using back-propagation. The proposed model is comprised of three networks, namely i) a feature extractor, ii) a functional connectivity network, and iii) a classification network.

Our findings highlight that functional connectivity serves as an important biomarker towards classification of ADHD and the frontal lobe is altered the most in the case of ADHD. The frontal lobe is known to be associated with cognitive functions like attention, memory, planning and mood. Our findings of the frontal lobe anomalies in ADHD support findings of the earlier studies. Our results reveal that an end to-end trainable deep network incorporating functional connectivity yields higher detection rates.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Departments:	Doctoral Theses School of Science & Technology > School of Science & Technology Doctoral Theses School of Science & Technology > Computer Science

Export

Downloads