Improving deep learning approach for FMRI data

Abstract

Functional magnetic resonance imaging (fMRI) has become one means to understand the epicentre of the human nervous system, the brain. It represents intrinsic haemodynamic signals in high-dimensional data that linked to neuronal activities. Numerous studies such as statistical parametric analysis, multivariate pattern analysis and few machine learning techniques have demonstrated significant results that suggested the links between human activities and the haemodynamic signals and which sections of the brain that got activated. In this research, deep learning, a recently discovered method that broke many benchmark records in areas such as object and speech recognition, is used. One of the significant advantages of deep learning is that it avoids the problem of labourintensive work, such as feature extractions. With that, various deep learning algorithms were studied and experimented. Contribution to the knowledge sphere in this thesis revolves around the data division for deep learning approach to classify high-dimensional fMRI when a rather low volume of data was adopted in the classification approach. High dimensionality and low signal-to-noise-ratio (SNR) are the biggest challenges in fMRI classification. Using a single centre slice of data reduces the anatomical variability dependence and curse of dimensionality degree. First, a minimal preprocessing stage is proposed for the two types dataset compilations; randomised and separated validation data of 1029 control fMRI individual subjects data. Convolutional neural network (CNN), studied and chosen deep learning method, has been assessed under three aspects: convolutional layers size; feature map sizes selection, and inception model blocks insertion. These aspects are few of many uncertainties in CNN modelling. The minimal preprocessing stage was proposed as opposed to lengthy conventional methods. Division of data shows the capabilities of deep learning to overfit the classification algorithm, though many adjustments were included. Besides, the model training step in processing stage formulates the problem as a single optimisation problem in which all the components of the model share a similar goal. It is an end-to-end deep learning algorithm reliability testing. This research requires very demanding computational capabilities with any increase in data volume. As a result, high accuracy was acquired with tested CNN models but inversely proportional for validation data accuracy when separated validation set was used. Although this research is designed for one slice of 3D fMRI data, an impressive set of computation resource such as a high-performance computer with stacked of dedicated graphic cards may have the ability to analyse a much higher volume of the whole-brain fMRI data. As a conclusion, this research shows that deep learning is reliable for classification but has the tendency to overfit and overgeneralisation. This was suggested when higher validation loss acquired with low volume of high-dimensional fMRI data employment. The data division strategy proposed in this research for end-to-end deep learning solution should be one of the keypoints for the data processing model.