Novel feature enhancement techniques for neural networks and their biomedical applications

Abstract

Neural networks (NNs) have emerged as a convenient tool for researchers/engineers to address complicated problems in a data-drive approach due to its high degree of freedom. In the processing/recognition of biological data such as computed tomography (CT), electroencephalogram (EEG) and functional Magnetic Resonance Imaging (fMRI), handcrafted features are conventionally employed, which may limit the ultimate performance. This thesis aims to develop feature enhancement techniques for NNs for these biomedical applications for better performance and resilient to discrepancy in multi-site data as well as possible label uncertainty. Particularly, we shall consider both stationary CT data for lung nodule detection/classification as well as time series data such as EEG and fMRI for classification of neuro-disorder such as Schizophrenia and Autism spectrum disorder (ASD). First, we propose an end-to-end NN for simultaneous detection, classification, and segmentation of lung nodules from 3-D CT images under label uncertainty. Our approach utilizes the features of the detection subnetwork to enhance the features and performance of the classification network. Both the nodule detection and classification subnetworks adopt a 3-D encoder-decoder architecture for better exploration of the 3-D data. Particularly, the classification subnetwork utilizes the multiscale nodule-specific features and the enhanced features from the detection sub-network for boosting classification performance. The latter serves as valuable prior information for directly optimizing the complicated 3-D classification network to better distinguish suspicious nodules from other tissues compared with direct backpropagation from the decoder. Experimental results show that the proposed nodule approach compares favorably to state-of-the-art algorithms. Secondly, the feature enhancing problem in learning multi-site data under possible dataset discrepancy is addressed. A novel adversarial-learning based multi-domain classification method is proposed for recognizing neuro-disorder using functional connectivity in fMRI. The system consists of a recognizer and a discriminator, which is trained to classify the data domain. Adversarial learning aims to extract dataset resilient features such that the discriminator can no longer distinguish the data domain. Our framework considers overall discrepancy rather than discrepancy between the adapted domain and the target domain as in conventional domain adaption approach. It also utilizes the attention mechanism to identify the relative importance of the instantaneous and lagged correlation as connectivity features. Experiments show that the approach yields better prediction performance than conventional approaches by exploring multi-domain discrepancy. Moreover, our framework can be more efficient than the domain adaptation method as it trains a single model to predict multi-site samples. Lastly, a Channel-Frequency Group Attention Network (CF-GAN) for feature enhancement in recognition of motor-imagery brain-computer interface and schizophrenia based on multi-channel EEGs is proposed. It explores similar characteristics in the channel-frequency domains of the EEG signals to simplify the network structure and reduce overfitting over typical attention approaches. An Expectation-Maximization (EM)–based algorithm is proposed to iteratively optimize the parameters of the GAN and the remaining prediction network. Moreover, sparsity promoting regularization is incorporated to reduce the electrodes needed, which leads to lower hardware complexity with comparable performance. Experimental results show that the proposed approach achieves state-of-the-art recognition performances in both Brain-Computer-Interface (BCI) and schizophrenia diagnosis.