Research On Key Issues In High Precision TDCS Brain Stimulation System

Transcranial direct current stimulation(tDCS)is a type of noninvasive transcranial electrical brain stimulation.Up to now,a majority of tDCS studies still used a large electrode pad to perform brain stimulation.Although this kind of stimulation mode could stimulate the target brain region,a large area of cortex around the target were also stimulated.Thus,the implementation of high precision tDCS brain stimulation can not only better help researchers to work on the tDCS study,but also promote the application of tDCS.Focus on key issues of the implementation of high precision tDCS brain stimulation,in this paper,we firstly researched on the localization problem of functional and structural target brain region using functional magnetic resonance imaging(fMRI)and structural MRI data.Secondly,based on 10/10 electroencephalography(EEG)electrode distribution,we researched on the current optimization problem on each tDCS electrode using dense array tDCS stimulation mode,which providing the technical support for the implementation of high precision tDCS stimulation in multi-brain region.Finally,we researched on the auto sleep stage classification problem using whole night EEG data of Sleep Heart Health Study(SHHS)dataset in National Sleep Research Resource public database,which providing the technical support for the online stimulation of tDCS during sleep.Specifically,the achievements of this paper can be summarized as follow.Firstly,we researched on the localization problem of functional target brain region using lifelong premature ejaculation(LPE)patients’ fMRI dataset.We firstly extracted 4005 functional connectivity(i.e.,features)from the fMRI data for each LPE patient and healthy control(HC)based on automated anatomical labeling(AAL)brain atlas.Next,we performed the feature selection step using cross validation based absolute shrinkage and selection operator(CV-LASSO)method.Finally,a support vector machine classification model was constructed to identify LPE patients from HCs using selected features.Results shown that several important features from 4005 features were successfully selected by the proposed CV-LASSO method.The constructed SVM model using these selected features could distinguish LPE patients from healthy controls with an accuracy of 0.8490±0.1401.In these selected features,5 features didn’t change with the different random grouping(100 times 10-folds CV).The 10 specific brain regions corresponding to these 5 features were all mentioned that had some relationship with LPE mechanism in previous LPE studies.Thus,these 10 brain regions can be identified as the functional target brain regions for the tDCS treatment,regulate and control or study in LPE.Secondly,we researched on the localization problem of structural target brain region using nasopharyngeal carcinoma(NPC)patients’ high-definition T2-weighted MRI dataset.We cascaded several deep learning networks and studied the performance of this cascaded network on NPC segmentation.Results shown that,the more the number of cascaded network,the better performance of NPC segmentation would be achieved.However,the performance would reach saturation with the number of cascaded network increased.Additionally,the simpler the structure of a network used for cascaded,the more performance gain will be gotten.In all of results,the model cascaded with three AlexNet achieved the best performance on NPC segmentation,with a mean dice similarity coefficient of 0.8335 in testing group(77 NPC patients).Thus,the exacted tumor from NPC patients’ MRI data using this model can be identified as the structural target region for the study and treatment of NPC.Thirdly,we researched on the precision stimulation problem of Dense array tDCS stimulation mode in multi-targets brain region case.We firstly achieved the high precision stimulation of Dense array tDCS in single target brain region case using minimum least squares(MLS)and maximum electrical field strength(ME)optimization scheme.Next,after improving and expanding these two scheme,we proposed weighted MLS and weighted ME optimization scheme and studied their stimulation performance in multi-targets brain region case.Results shown that our proposed weighted MLS and weighted ME scheme could not only achieve the high precision stimulation in multi-targets brain region case,which the MLS and the ME scheme could not achieve,but also precisely distribute stimulation energy into each target brain region.Finally,we researched on the sleep EEG data based sleep stage classification problem using Sleep Heart Health Study dataset.Firstly,we continuously divided the whole night EEG signal into several small windows,each window containing 30 s EEG signal.Then,we calculated the time frequency spectra in each window using short-time Fourier transform algorithm.Next,with time frequency spectra of single time point(single 30 s window)as input,we studied the performance of convolutional neural network(CNN)and long short term memory(LSTM)network on sleep stage classification,respectively.Finally,with time frequency spectra of multi-time points as input,we further studied the sleep stage classification performance of LSTM network.Results shown that,considering the time relationship in time frequency spectra of single time point,the performance of LSTM network on sleep stage classification was improved to some extent when comparing with traditional CNN network.When considering the time relationship in time frequency spectra of multi-time points,the sleep stage classification performance of LSTM network was further improved and reached the saturation when considering three time points,with an accuracy of 87.4% and a Cohen’s kappa coefficient of 0.8216.Thus,with time frequency spectra of multi-time points as input,the LSTM network could well perform the sleep stage classification task.In summary,four issues in high precision tDCS brain stimulation had been studied in this paper: 1+2)functional target location issue and structural target location issue(target location);3)multi-target high precision tDCS stimulation issue(high precision stimulation);4)auto sleep stage classification issue(stimulation timing identification during sleep).The proposed CV-LASSO method and cascaded deep learning model could locate the functional and structural target from fMRI and high-definition MRI images with a better precision,respectively.The proposed weighted MLS and weighted ME optimization schemes could achieve a high precision stimulation in multi-target brain regions.The proposed LSTM based sleep stage classification method could achieve a precise stimulation timing identification during sleep.