Background: The automatic classification of electrocardiogram (ECG) data using a convolutional neural network (CNN) model has been practiced earlier, but there are only a few studies on a 12-lead ECG dataset with various class labels. A large amount of ECG data is stored in hospital information systems in Japan, and this data can be used for machine learning. However, each sample in the data is mostly recorded for ten seconds and labelled with the corresponding abnormal classes, not for each lead or waveform, but for the entire 12-lead dataset. Therefore, the one-shot screening method using 2-D images of superimposed PQRST waveforms can be a solution in the given condition that all waveforms in a sample within a certain duration must be processed simultaneously.
Objective: We propose the one-shot screening method with different types of 2-D images of superimposed PQRST waveforms using CNN.
Methods: CNN and ensemble learning were applied to the ECG dataset, which contains over 9,000 samples with two classes, normal and abnormal, consolidated from 130 abnormal class labels for binary classification. We prepared three types of ECG images that were different in the manner in which they superimposed the PQRST waveforms of a single heartbeat: left-aligned, right-aligned, and centered. We compared the results of the three different images and analyzed false negative patterns to ascertain the characteristics of different types of 2D-CNN.
Results: The accuracy of all the frameworks were found to be above 0.867. The framework with the centered ECG images achieved the highest accuracy of 0.938 among the three. The listed abnormal classes with a high false negative ratio differed on the basis of the type of model.
Conclusion: The model with centered images showed the best score with the application of the one-shot method; however, the error analysis demonstrated that the characteristics of these models are varied.
AI technique to detemine the level of cognitive/intelligent level.
Background: The fiber-tracking based on diffusion
magnetic resonance imaging technique is one of the methods that indirectly
examine the structural information about the cerebral white matter in vivo.
Conventionally, fiber-tracking approaches are accompanied with manual settings
of the starting regions of fiber tracking, which is usually very time-consuming
and unsuitable for large-scale data analysis.
order to clarify the physical fiber connections associated with
disease-specific neural circuits for various neuropsychiatric disorders, we
proposed an automated diffusion magnetic resonance imaging based whole-brain
fiber-tracking method combined with an atlas separating 54 regions of the white
Methods: The propose method automatically set the
fiber-tracking starting plane in each parcel and rotated along the running
direction of fibers as well so that it enabled swift, objective, and
reproducible analysis. The method was verified with real diffusion magnetic
resonance imaging data recorded from three healthy volunteers.
Results: The mean fiber direction for each parcel was confirmed to fit the anatomical
configuration. The fiber-tracking streamlines were confirmed to run along the
first eigenvector of diffusion tensor and to terminate according to the pre-set
termination condition. The major fiber tracts of the 54 white matter parcels
were relevantly reconstructed.
Conclusion: The results
demonstrate the feasibility of the proposed automated whole-brain atlas-based
fiber-tracking method for investigating white matter disruptions associated
with neuropsychiatric disorders especially in large-scale datasets.