Motor Imagery Classification for Brain Computer Interface Using Deep Convolutional Neural Networks and Mixup Augmentation

Goal: Building a DL model that can be trained on small EEG training set of a single subject presents an interesting challenge that this work is trying to address. In particular, this study is trying to avoid the need for long EEG data collection sessions, and without combining multiple subjects training datasets, which has a detrimental effect on the classification performance due to the inter-individual variability among subjects. Methods: A customized Convolutional Neural Network with mixup augmentation was trained with [Formula: see text]120 EEG trials for only one subject per model. Results: Modified ResNet18 and DenseNet121 models with mixup augmentation achieved 0.920 (95% Confidence Interval: 0.908, 0.933) and 0.933 (95% Confidence Interval: 0.922, 0.945) classification accuracy, respectively. Conclusions: We show that the designed classifiers resulted in a higher classification performance in comparison to other DL classifiers of previous studies on the same dataset, despite the limited training dataset used in this work.

Blind-Spot Collision Detection System for Commercial Vehicles Using Multi Deep CNN Architecture

Buses and heavy vehicles have more blind spots compared to cars and other road vehicles due to their large sizes. Therefore, accidents caused by these heavy vehicles are more fatal and result in severe injuries to other road users. These possible blind-spot collisions can be identified early using vision-based object detection approaches. Yet, the existing state-of-the-art vision-based object detection models rely heavily on a single feature descriptor for making decisions. In this research, the design of two convolutional neural networks (CNNs) based on high-level feature descriptors and their integration with faster R-CNN is proposed to detect blind-spot collisions for heavy vehicles. Moreover, a fusion approach is proposed to integrate two pre-trained networks (i.e., Resnet 50 and Resnet 101) for extracting high level features for blind-spot vehicle detection. The fusion of features significantly improves the performance of faster R-CNN and outperformed the existing state-of-the-art methods. Both approaches are validated on a self-recorded blind-spot vehicle detection dataset for buses and an online LISA dataset for vehicle detection. For both proposed approaches, a false detection rate (FDR) of 3.05% and 3.49% are obtained for the self recorded dataset, making these approaches suitable for real time applications.

Effective Connectivity for Decoding Electroencephalographic Motor Imagery Using a Probabilistic Neural Network

Motor imagery (MI)-based brain-computer interfaces have gained much attention in the last few years. They provide the ability to control external devices, such as prosthetic arms and wheelchairs, by using brain activities. Several researchers have reported the inter-communication of multiple brain regions during motor tasks, thus making it difficult to isolate one or two brain regions in which motor activities take place. Therefore, a deeper understanding of the brain's neural patterns is important for BCI in order to provide more useful and insightful features. Thus, brain connectivity provides a promising approach to solving the stated shortcomings by considering inter-channel/region relationships during motor imagination. This study used effective connectivity in the brain in terms of the partial directed coherence (PDC) and directed transfer function (DTF) as intensively unconventional feature sets for motor imagery (MI) classification. MANOVA-based analysis was performed to identify statistically significant connectivity pairs. Furthermore, the study sought to predict MI patterns by using four classification algorithms-an SVM, KNN, decision tree, and probabilistic neural network. The study provides a comparative analysis of all of the classification methods using two-class MI data extracted from the PhysioNet EEG database. The proposed techniques based on a probabilistic neural network (PNN) as a classifier and PDC as a feature set outperformed the other classification and feature extraction techniques with a superior classification accuracy and a lower error rate. The research findings indicate that when the PDC was used as a feature set, the PNN attained the greatest overall average accuracy of 98.65%, whereas the same classifier was used to attain the greatest accuracy of 82.81% with the DTF. This study validates the activation of multiple brain regions during a motor task by achieving better classification outcomes through brain connectivity as compared to conventional features. Since the PDC outperformed the DTF as a feature set with its superior classification accuracy and low error rate, it has great potential for application in MI-based brain-computer interfaces.

Deep learning architectures analysis for age-related macular degeneration segmentation on optical coherence tomography scans

BACKGROUND AND OBJECTIVES

Aged people usually are more to be diagnosed with retinal diseases in developed countries. Retinal capillaries leakage into the retina swells and causes an acute vision loss, which is called age-related macular degeneration (AMD). The disease can not be adequately diagnosed solely using fundus images as depth information is not available. The variations in retina volume assist in monitoring ophthalmological abnormalities. Therefore, high-fidelity AMD segmentation in optical coherence tomography (OCT) imaging modality has raised the attention of researchers as well as those of the medical doctors. Many methods across the years encompassing machine learning approaches and convolutional neural networks (CNN) strategies have been proposed for object detection and image segmentation.

METHODS

In this paper, we analyze four wide-spread deep learning models designed for the segmentation of three retinal fluids outputting dense predictions in the RETOUCH challenge data. We aim to demonstrate how a patch-based approach could push the performance for each method. Besides, we also evaluate the methods using the OPTIMA challenge dataset for generalizing network performance. The analysis is driven into two sections: the comparison between the four approaches and the significance of patching the images.

RESULTS

The performance of networks trained on the RETOUCH dataset is higher than human performance. The analysis further generalized the performance of the best network obtained by fine-tuning it and achieved a mean Dice similarity coefficient (DSC) of 0.85. Out of the three types of fluids, intraretinal fluid (IRF) is more recognized, and the highest DSC value of 0.922 is achieved using Spectralis dataset. Additionally, the highest average DSC score is 0.84, which is achieved by PaDeeplabv3+ model using Cirrus dataset.

CONCLUSIONS

The proposed method segments the three fluids in the retina with high DSC value. Fine-tuning the networks trained on the RETOUCH dataset makes the network perform better and faster than training from scratch. Enriching the networks with inputting a variety of shapes by extracting patches helped to segment the fluids better than using a full image.

Single-Trial Visual Evoked Potential Extraction Using Partial Least-Squares-Based Approach

A single-trial extraction of a visual evoked potential (VEP) signal based on the partial least-squares (PLS) regression method has been proposed in this paper. This paper has focused on the extraction and estimation of the latencies of P100, P200, P300, N75, and N135 in the artificial electroencephalograph (EEG) signal. The real EEG signal obtained from the hospital was only concentrated on the P100. The performance of the PLS has been evaluated mainly on the basis of latency error rate of the peaks for the artificial EEG signal, and the mean peak detection and standard deviation for the real EEG signal. The simulation results show that the proposed PLS algorithm is capable of reconstructing the EEG signal into its desired shape of the ideal VEP. For P100, the proposed PLS algorithm is able to provide comparable results to the generalized eigenvalue decomposition (GEVD) algorithm, which alters (prewhitens) the EEG input signal using the prestimulation EEG signal. It has also shown better performance for later peaks (P200 and P300). The PLS outperformed not only in positive peaks but also in N75. In P100, the PLS was comparable with the GEVD although N135 was better estimated by GEVD. The proposed PLS algorithm is comparable to GEVD given that PLS does not alter the EEG input signal. The PLS algorithm gives the best estimate to multitrial ensemble averaging. This research offers benefits such as avoiding patient's fatigue during VEP test measurement in the hospital, in BCI applications and in EEG-fMRI integration.

Single-trial subspace-based approach for VEP extraction

A signal subspace approach for extracting visual evoked potentials (VEPs) from the background electroencephalogram (EEG) colored noise without the need for a prewhitening stage is proposed. Linear estimation of the clean signal is performed by minimizing signal distortion while maintaining the residual noise energy below some given threshold. The generalized eigendecomposition of the covariance matrices of a VEP signal and brain background EEG noise is used to transform them jointly to diagonal matrices. The generalized subspace is then decomposed into signal subspace and noise subspace. Enhancement is performed by nulling the components in the noise subspace and retaining the components in the signal subspace. The performance of the proposed algorithm is tested with simulated and real data, and compared with the recently proposed signal subspace techniques. With the simulated data, the algorithms are used to estimate the latencies of P(100), P(200), and P(300) of VEP signals corrupted by additive colored noise at different values of SNR. With the real data, the VEP signals are collected at Selayang Hospital, Kuala Lumpur, Malaysia, and the capability of the proposed algorithm in detecting the latency of P(100) is obtained and compared with other subspace techniques. The ensemble averaging technique is used as a baseline for this comparison. The results indicated significant improvement by the proposed technique in terms of better accuracy and less failure rate.

A generalized subspace approach for estimating visual evoked potentials

A "single-trial" signal subspace approach for extracting visual evoked potential (VEP) from the ongoing 'colored' electroencephalogram (EEG) noise is proposed. The algorithm applies the generalized eigendecomposition on the covariance matrices of the VEP and noise to transform them jointly into diagonal matrices in order to avoid a pre-whitening stage. The proposed generalized subspace approach (GSA) decomposes the corrupted VEP space into a signal subspace and noise subspace. Enhancement is achieved by removing the noise subspace and estimating the clean VEPs only from the signal subspace. The validity and effectiveness of the proposed GSA scheme in estimating the latencies of P100's (used in objective assessment of visual pathways) are evaluated using real data collected from Selayang Hospital in Kuala Lumpur. The performance of GSA is compared with the recently proposed single-trial technique called the Third Order Correlation (TOC).