Schizophrenia is a severe neuropsychiatric disease whose diagnosis, unfortunately, lacks an objective diagnostic tool supporting a thorough psychiatric examination of the patient. We took advantage of today's computational abilities, structural magnetic resonance imaging, and modern machine learning methods, such as stacked autoencoders (SAE) and 3D convolutional neural networks (3D CNN), to teach them to classify 52 patients with schizophrenia and 52 healthy controls. The main aim of this study was to explore whether complex feature extraction methods can help improve the accuracy of deep learning-based classifiers compared to minimally preprocessed data. Our experiments employed three commonly used preprocessing steps to extract three different feature types. They included voxel-based morphometry, deformation-based morphometry, and simple spatial normalization of brain tissue. In addition to classifier models, features and their combination, other model parameters such as network depth, number of neurons, number of convolutional filters, and input data size were also investigated. Autoencoders were trained on feature pools of 1000 and 5000 voxels selected by Mann-Whitney tests, and 3D CNNs were trained on whole images. The most successful model architecture (autoencoders) achieved the highest average accuracy of 69.62% (sensitivity 68.85%, specificity 70.38%). The results of all experiments were statistically compared (the Mann-Whitney test). In conclusion, SAE outperformed 3D CNN, while preprocessing using VBM helped SAE improve the results.
- Publication type
- Journal Article MeSH
Manual visual review, annotation and categorization of electroencephalography (EEG) is a time-consuming task that is often associated with human bias and requires trained electrophysiology experts with specific domain knowledge. This challenge is now compounded by development of measurement technologies and devices allowing large-scale heterogeneous, multi-channel recordings spanning multiple brain regions over days, weeks. Currently, supervised deep-learning techniques were shown to be an effective tool for analyzing big data sets, including EEG. However, the most significant caveat in training the supervised deep-learning models in a clinical research setting is the lack of adequate gold-standard annotations created by electrophysiology experts. Here, we propose a semi-supervised machine learning technique that utilizes deep-learning methods with a minimal amount of gold-standard labels. The method utilizes a temporal autoencoder for dimensionality reduction and a small number of the expert-provided gold-standard labels used for kernel density estimating (KDE) maps. We used data from electrophysiological intracranial EEG (iEEG) recordings acquired in two hospitals with different recording systems across 39 patients to validate the method. The method achieved iEEG classification (Pathologic vs. Normal vs. Artifacts) results with an area under the receiver operating characteristic (AUROC) scores of 0.862 ± 0.037, 0.879 ± 0.042, and area under the precision-recall curve (AUPRC) scores of 0.740 ± 0.740, 0.714 ± 0.042. This demonstrates that semi-supervised methods can provide acceptable results while requiring only 100 gold-standard data samples in each classification category. Subsequently, we deployed the technique to 12 novel patients in a pseudo-prospective framework for detecting Interictal epileptiform discharges (IEDs). We show that the proposed temporal autoencoder was able to generalize to novel patients while achieving AUROC of 0.877 ± 0.067 and AUPRC of 0.705 ± 0.154.
BACKGROUND: This study aims to evaluate the feasibility of generating pseudo-normal single photon emission computed tomography (SPECT) data from potentially abnormal images. These pseudo-normal images are primarily intended for use in an on-the-fly data harmonization technique. MATERIAL AND METHODS: The methodology was tested on brain SPECT with [123I]Ioflupane. The proposed model for generating a pseudo-normal image was based on a variational autoencoder (VAE) designed to process 2D sinograms of the brain [123I]-FP-CIT SPECT, potentially exhibiting abnormal uptake. The model aimed to predict SPECT sinograms with corresponding normal uptake. Training, validation, and testing datasets were created by SPECT simulator from 45 brain masks segmented from real patient's magnetic resonance (MR) scans, using various uptake levels. The training and validation datasets each comprised 612 and 360 samples, respectively, drawn from 36 brain masks. The testing dataset contained 153 samples based on 9 brain masks. VAE performance was evaluated through brain dimensions, Dice similarity coefficient (DSC) and specific binding ratio. RESULTS: Mean DSC was 80% for left basal ganglia and 84% for right basal ganglia. The proposed VAE demonstrated excellent consistency in predicting basal ganglia shape, with a coefficient of variation of DSC being less than 1.1%. CONCLUSIONS: The study demonstrates that VAE can effectively estimate an individualized pseudo-normal distribution of the radiotracer [123I]-FP-CIT SPECT from abnormal SPECT images. The main limitations of this preliminary research are the limited availability of real brain MR data, used as input for the SPECT data simulator, and the simplified simulation setup employed to create the synthetic dataset.
- MeSH
- Tomography, Emission-Computed, Single-Photon * MeSH
- Humans MeSH
- Brain * diagnostic imaging MeSH
- Image Processing, Computer-Assisted * MeSH
- Check Tag
- Humans MeSH
- Publication type
- Journal Article MeSH
The current population worldwide extensively uses social media to share thoughts, societal issues, and personal concerns. Social media can be viewed as an intelligent platform that can be augmented with a capability to analyze and predict various issues such as business needs, environmental needs, election trends (polls), governmental needs, etc. This has motivated us to initiate a comprehensive search of the COVID-19 pandemic-related views and opinions amongst the population on Twitter. The basic training data have been collected from Twitter posts. On this basis, we have developed research involving ensemble deep learning techniques to reach a better prediction of the future evolutions of views in Twitter when compared to previous works that do the same. First, feature extraction is performed through an N-gram stacked autoencoder supervised learning algorithm. The extracted features are then involved in a classification and prediction involving an ensemble fusion scheme of selected machine learning techniques such as decision tree (DT), support vector machine (SVM), random forest (RF), and K-nearest neighbour (KNN). all individual results are combined/fused for a better prediction by using both mean and mode techniques. Our proposed scheme of an N-gram stacked encoder integrated in an ensemble machine learning scheme outperforms all the other existing competing techniques such unigram autoencoder, bigram autoencoder, etc. Our experimental results have been obtained from a comprehensive evaluation involving a dataset extracted from open-source data available from Twitter that were filtered by using the keywords "covid", "covid19", "coronavirus", "covid-19", "sarscov2", and "covid_19".
- MeSH
- COVID-19 * MeSH
- Humans MeSH
- Pandemics MeSH
- SARS-CoV-2 MeSH
- Social Media * MeSH
- Social Networking MeSH
- Machine Learning MeSH
- Check Tag
- Humans MeSH
- Publication type
- Journal Article MeSH
Electroencephalography (EEG) experiments typically generate vast amounts of data due to the high sampling rates and the use of multiple electrodes to capture brain activity. Consequently, storing and transmitting these large datasets is challenging, necessitating the creation of specialized compression techniques tailored to this data type. This study proposes one such method, which at its core uses an artificial neural network (specifically a convolutional autoencoder) to learn the latent representations of modelled EEG signals to perform lossy compression, which gets further improved with lossless corrections based on the user-defined threshold for the maximum tolerable amplitude loss, resulting in a flexible near-lossless compression scheme. To test the viability of our approach, a case study was performed on the 256-channel binocular rivalry dataset, which also describes mostly data-specific statistical analyses and preprocessing steps. Compression results, evaluation metrics, and comparisons with baseline general compression methods suggest that the proposed method can achieve substantial compression results and speed, making it one of the potential research topics for follow-up studies.
- MeSH
- Adult MeSH
- Electroencephalography * methods MeSH
- Data Compression * methods MeSH
- Humans MeSH
- Neural Networks, Computer * MeSH
- Signal Processing, Computer-Assisted MeSH
- Check Tag
- Adult MeSH
- Humans MeSH
- Male MeSH
- Female MeSH
- Publication type
- Journal Article MeSH
BACKGROUND: Recent advances in data-driven computational approaches have been helpful in devising tools to objectively diagnose psychiatric disorders. However, current machine learning studies limited to small homogeneous samples, different methodologies, and different imaging collection protocols, limit the ability to directly compare and generalize their results. Here we aimed to classify individuals with PTSD versus controls and assess the generalizability using a large heterogeneous brain datasets from the ENIGMA-PGC PTSD Working group. METHODS: We analyzed brain MRI data from 3,477 structural-MRI; 2,495 resting state-fMRI; and 1,952 diffusion-MRI. First, we identified the brain features that best distinguish individuals with PTSD from controls using traditional machine learning methods. Second, we assessed the utility of the denoising variational autoencoder (DVAE) and evaluated its classification performance. Third, we assessed the generalizability and reproducibility of both models using leave-one-site-out cross-validation procedure for each modality. RESULTS: We found lower performance in classifying PTSD vs. controls with data from over 20 sites (60 % test AUC for s-MRI, 59 % for rs-fMRI and 56 % for d-MRI), as compared to other studies run on single-site data. The performance increased when classifying PTSD from HC without trauma history in each modality (75 % AUC). The classification performance remained intact when applying the DVAE framework, which reduced the number of features. Finally, we found that the DVAE framework achieved better generalization to unseen datasets compared with the traditional machine learning frameworks, albeit performance was slightly above chance. CONCLUSION: These results have the potential to provide a baseline classification performance for PTSD when using large scale neuroimaging datasets. Our findings show that the control group used can heavily affect classification performance. The DVAE framework provided better generalizability for the multi-site data. This may be more significant in clinical practice since the neuroimaging-based diagnostic DVAE classification models are much less site-specific, rendering them more generalizable.
