Introducing a Novel Layer in Convolutional Neural Network for Automatic Identification of Diabetic Retinopathy

Convolutional neural networks have been widely used for identifying diabetic retinopathy on color fundus images. For such application, we proposed a novel framework for the convolutional neural network architecture by embedding a preprocessing layer followed by the first convolutional layer to increase the performance of the convolutional neural network classifier. Two image enhancement techniques i.e. 1- Contrast Enhancement 2- Contrast-limited adaptive histogram equalization were separately embedded in the proposed layer and the results were compared. For identification of exudates, hemorrhages and microaneurysms, the proposed framework achieved the total accuracy of 87.6%, and 83.9% for the contrast enhancement and contrast-limited adaptive histogram equalization layers, respectively. However, the total accuracy of the convolutional neural network alone without the prreprocessing layer was found to be 81.4%. Consequently, the new convolutional neural network architecture with the proposed preprocessing layer improved the performance of convolutional neural network.

Age-related rarefaction in retinal vasculature is not linear

The fractal dimension is a global measure of complexity and is useful for quantifying anatomical structures, including the retinal vascular network. A previous study found a linear declining trend with aging on the retinal vascular fractal dimension (DF); however, it was limited to the older population (49 years and older). This study aimed to investigate the possible models of the fractal dimension changes from young to old subjects (10–73 years). A total of 215 right-eye retinal samples, including those of 119 (55%) women and 96 (45%) men, were selected. The retinal vessels were segmented using computer-assisted software, and non-vessel fragments were deleted. The fractal dimension was measured based on the log–log plot of the number of grids versus the size. The retinal vascular DF was analyzed to determine changes with increasing age. Finally, the data were fitted to three polynomial models. All three models are statistically significant (Linear: R(2) = 0.1270, 213 d.f., p < 0.001, Quadratic: R(2) = 0.1536, 212 d.f., p < 0.001, Cubic: R(2) = 0.1529, 211 d.f., p < 0.001). The quadratic regression is significantly better than the linear regression (p < 0.001); however, the increase in R(2) from the quadratic model to the cubic model is not significant (p = 0.97). These results suggest that the decreasing trend of the fractal dimension associated with aging is better explained by the quadratic model than by the linear and cubic models in a sample with a broader age spectrum.

Age-related rarefaction in retinal vasculature is not linear

The fractal dimension is a global measure of complexity and is useful for quantifying anatomical structures, including the retinal vascular network. A previous study found a linear declining trend with aging on the retinal vascular fractal dimension (D_F); however, it was limited to the older population (49 years and older). This study aimed to investigate the possible models of the fractal dimension changes from young to old subjects (10-73 years). A total of 215 right-eye retinal samples, including those of 119 (55%) women and 96 (45%) men, were selected. The retinal vessels were segmented using computer-assisted software, and non-vessel fragments were deleted. The fractal dimension was measured based on the log-log plot of the number of grids versus the size. The retinal vascular D_F was analyzed to determine changes with increasing age. Finally, the data were fitted to three polynomial models. All three models are statistically significant (Linear: R² = 0.1270, 213 d.f., p < 0.001, Quadratic: R² = 0.1536, 212 d.f., p < 0.001, Cubic: R² = 0.1529, 211 d.f., p < 0.001). The quadratic regression is significantly better than the linear regression (p < 0.001); however, the increase in R² from the quadratic model to the cubic model is not significant (p = 0.97). These results suggest that the decreasing trend of the fractal dimension associated with aging is better explained by the quadratic model than by the linear and cubic models in a sample with a broader age spectrum.

Robust methodology for fractal analysis of the retinal vasculature

We have developed a robust method to perform retinal vascular fractal analysis from digital retina images. The technique preprocesses the green channel retina images with Gabor wavelet transforms to enhance the retinal images. Fourier Fractal dimension is computed on these preprocessed images and does not require any segmentation of the vessels. This novel technique requires human input only at a single step; the allocation of the optic disk center. We have tested this technique on 380 retina images from healthy individuals aged 50+ years, randomly selected from the Blue Mountains Eye Study population. To assess its reliability in assessing retinal vascular fractals from different allocation of optic center, we performed pair-wise Pearson correlation between the fractal dimension estimates with 100 simulated region of interest for each of the 380 images. There was Gaussian distribution variation in the optic center allocation in each simulation. The resulting mean correlation coefficient (standard deviation) was 0.93 (0.005). The repeatability of this method was found to be better than the earlier box-counting method. Using this method to assess retinal vascular fractals, we have also confirmed a reduction in the retinal vasculature complexity with aging, consistent with observations from other human organ systems.

A machine learning based method for classification of fractal features of forearm sEMG using Twin Support vector machines

Classification of surface electromyogram (sEMG) signal is important for various applications such as prosthetic control and human computer interface. Surface EMG provides a better insight into the strength of muscle contraction which can be used as control signal for different applications. Due to the various interference between different muscle activities, it is difficult to identify movements using sEMG during low-level flexions. A new set of fractal features - fractal dimension and Maximum fractal length of sEMG has been previously reported by the authors. These features measure the complexity and strength of the muscle contraction during the low-level finger flexions. In order to classify and identify the low-level finger flexions using these features based on the fractal properties, a recently developed machine learning based classifier, Twin Support vector machines (TSVM) has been proposed. TSVM works on basic learning methodology and solves the classification tasks as two SVMs for each classes. This paper reports the novel method on the machine learning based classification of fractal features of sEMG using the Twin Support vector machines. The training and testing was performed using two different kernel functions - Linear and Radial Basis Function (RBF).

Surface EMG based hand gesture identification using semi blind ICA: validation of ICA matrix analysis

Surface electromyogram (sEMG) has numerous applications. It has been widely used in various biosignal and neuro rehabilitation applications. There is an urgent need for establishing a simple yet robust system that can be used to identify subtle complex hand actions and gestures for control of prosthesis and other computer assisted devices. Earlier work to identify the hand actions and gestures based on sEMG suffers from limitation that these are suitable for gross actions where there is only one prime-mover muscle involved and not suitable for small subtle and complex muscle contraction. This paper presents the hand gesture identification using sEMG decomposed using semi-blind independent component analysis combined with neural network based classifier. The aim was to provide reliable and natural control for rehabilitation and human computer interaction applications. We have proposed a model based approach where the hand muscle anatomy is known. The system was tested on 5 subjects and with experiments repeated on different days. The system was compared with raw sEMG as used by other researchers. The system is able to classify the different hand actions 100%. In comparison, the classification of the traditional ICA and raw sEMG for the same experiments and similar features was a poor 65% and 60% respectively. This research demonstrates that sEMG can be decomposed to the individual muscle activities using semi-blind ICA. The muscle activity after decomposition can be used to accurately identify small and subtle hand actions and gestures. Finally the ICA source separation was validated with mixing matrix analysis.

Strategies to identify changes in SEMG due to muscle fatigue during cycling

Detection, quantification and analysis of muscle fatigue are crucial in occupational/rehabilitation and sporting settings. Sports organizations such as the Australian Institute of Sports (AIS) currently monitor fatigue by a battery of tests including invasive techniques that require taking blood samples and/or muscle biopsies, the latter of which is highly invasive, painful, time consuming and expensive. SEMG is non-invasive monitoring of muscle activation and is an indication of localized muscle fatigue based on the observed shift of the power spectral density of the SEMG. But the success of SEMG based techniques is currently limited to isometric contraction and is not acceptable to the human movement community. This paper proposes and tests the use of spectral analysis of narrow windows of SEMG near the peak of a cyclic activity to identify the onset of muscle fatigue during cyclic activities. The results demonstrate a highly significant relationship of reduction of the median frequency with the onset of muscle fatigue. The paper also reports the validation of the SEMG study using biochemical analysis of muscle biopsy and blood tests and further verified using power output of the cycle and speed of pedalling.

Electrocardiogram removal from electromyogram of the muscles

Surface electromyogram (SEMG) of the lumbar back muscles is being used for determining posture disorders for people suffering from low back pain. But SEMG of the back has a strong electrocardiogram (ECG) artefact. Research was conducted to determine the difference in the SEMG before and after the removal of ECG artefact from the SEMG recording using gating, subtraction and multi-step independent component analysis (MICA). The paper reports results of experiments conducted on eleven subjects over two days. The results show that removal of ECG artefact from the raw signal can substantially alter the RMS of the signal demonstrating the need for careful filtering of the signal for analysis. The paper also reports of the success of MICA for removing the ECG artefact.

Relationship of magnitude of electromyogram of the lumbar muscles to static posture

This work reports research that investigated the relationship of the strength of contraction of different muscles of the lumbar back to maintain static posture. The paper reports the study of surface electromyogram of the muscles and uses scattering and neural networks on the strength of the EMG measured using root mean square (RMS). The signal is studied before and after the removal of electrocardiogram (ECG) artifact from the signal using a modified independent component analysis technique. The three dimensional scattering plots do not show any observable trends while the neural networks before the removal of ECG does not converge while after removal of ECG the neural network shows a high level of accuracy. The results demonstrate that there is a complex relation between the four EMG and posture and this relationship is revealed only after of ECG from EMG.

Limitations and applications of ICA for surface electromyogram

Surface electromyogram (SEMG) has numerous applications, but the presence of artefacts and noise, especially at low level of muscle activity make the recordings unreliable. Spectral and temporal overlap can make the removal of artefacts and noise, or separation of relevant signals from other bioelectric signals extremely difficult. Individual muscles may be considered as independent at the local level and this makes an argument for separating the signals using independent component analysis (ICA). In the recent past, due to the easy availability of ICA tools, numbers of researchers have attempted to use ICA for this application. This paper reports research conducted to evaluate the use of ICA for the separation of muscle activity and removal of the artefacts from SEMG. It discusses some of the conditions that could affect the reliability of the separation and evaluates issues related to the properties of the signals and number of sources. The paper also identifies the lack of suitable measure of quality of separation for bioelectric signals and it recommends and tests a more robust measure of separation. The paper also reports tests using Zibulevsky's technique of temporal plotting to identify number of independent sources in SEMG recordings. The theoretical analysis and experimental results demonstrate that ICA is suitable for SEMG signals. The results identify the unsuitability of ICA when the number of sources is greater than the number of recording channels. The results also demonstrate the limitations of such applications due to the inability of the system to identify the correct order and magnitude of the signals. The paper determines the suitability of the use of error measure using simulated mixing matrix and the estimated unmixing matrix as a means identifying the quality of separation of the output. The work demonstrates that even at extremely low level of muscle contraction, and with filtering using wavelets and band pass filters, it is not possible to get the data sparse enough to identify number of independent sources using Zibulevs.ky's technique.

Strategies to identify changes in SEMG due to muscle fatigue during cycling

Detection, quantification and analysis of muscle fatigue is crucial in occupational/rehabilitation and sporting settings. Sports organizations such as Australian Institute of Sports (AIS) currently monitor fatigue by a battery of tests including invasive techniques that require taking blood samples and/or muscle biopsies, the latter of which is highly invasive, painful, time consuming and expensive. SEMG is non-invasive monitoring of the muscle activation and is an indication of localized muscle fatigue based on the observed shift of the power spectral density of the SEMG. But the success of SEMG based techniques is currently limited to isometric contraction and is not acceptable to the human movement community. This paper proposes and tests a simple signal processing technique to identify the onset of muscle fatigue during cyclic activities of muscles such as VL and VM during cycling. Based on experiments conducted with 7 participants, using power output as a measure of fatigue, the technique is able to identify the muscle fatigue with 98% significance.

EEG coherence changes between right and left motor cortical areas during voluntary muscular contraction

It is known that movements of the right side of the body are controlled by the left motor cortex of the brain. The aim of this study is to evaluate the contribution of right motor cortex of the brain in the central motor control of right-sided muscle contraction. EEG/EEG coherence analysis has been used to determine the functional coupling between the right and left motor cortical areas in twenty normal volunteers, during maximum voluntary contraction (MVC) and 50% MVC of right Adductor Pollicis muscle (APM). It shows that the maximum mean coherence values were: 0.751 during MVC at 10 and 12 Hz, and 0.274 during 50% of MVC at 22 Hz. The minimum mean coherence values were: 0.716 during MVC at 48 and 50 Hz, and 0.242 during 50% MVC at 34 Hz. The high coherence values obtained during MVC, and to a lesser extent during 50% of MVC, could be attributed to the need of recruitment of both motor cortical areas during the decision phase of central motor control of voluntary muscular contraction. The "will" to perform maximum voluntary contraction could be a major factor, which contribute to the higher coherence values obtained during MVC than these associated with 50% of MVC.

Medical e-commerce for regional Australia

The residents of rural and regional Australia have less access to health care services than in capital cities. There is a reluctance of General Practitioners to practice in the country. New information technology and government initiatives are now addressing this problem. High bandwidth videoconferencing is now being routinely used to provide psychiatric consultations to areas without this service. But this (like many other implementations of telecommunication technologies to health) has resulted in loss of revenue to regional Australia while benefiting capital cities. Thus, the current implementation of telecommunication technology to health has resulted in loss of revenue of the regions while increasing the bias towards the cities. Further, the system is not economically viable and requires the Government to inject funds for the smooth operation of the system. This paper proposes the use of telecommunication technology for enabling the communities of regional Australia to access health facilities via physical and virtual clinics. The proposed technique is self supporting and is based in the country with the intent to prevent the drain of resources from regional Australia. The technique attempts to eradicate the problem at the root level by providing a business opportunity that is based in and to cater for the needs of the remote communities. The proposed system would provide health services by physical and virtual clinics and while serving the communities would be profit centres- and thus attracting doctors and other resources to the remote communities.

Wavelets for QRS detection

This paper examines the use of different wavelet functions for QRS complex detection in ECG. Wavelets provide time and frequency analysis simultaneously and offer flexibility with a number of wavelet functions with different properties available. This research has examined wavelet functions with different properties to determine the effects of orthogonality and time/frequency compactness of the wavelet on the ability to correctly detect the QRS. The error in detection (false negatives and positives) is the criterion for determining the efficacy of the wavelet function. The paper reports a significant reduction in error in detection of QRS complexes with mean error reduced to 0.75%. It also reports that wavelet functions that support symmetry and compactness provide better results.

Classification of dynamic multi-channel Electromyography by Neural Network

Muscles are responsible for movement of the limbs. Muscle contraction is accompanied by electrical activity that is measurable and is the Electromyography (EMG) recording. Due to the complex nature of the signal, detailed analysis and classification is often difficult, especially if the EMG relates to movement. This paper reports the research to determine features of the multi-channel EMG signal recording that correlate with the movement of the hand of the subject. Different processing techniques are reported. It demonstrates integral of the RMS of the signal correlates best with the movement.

The influence of Inter-Electrode Distance on the RMS of the SEMG signal

The aim of this research was to study the dependency of RMS of SEMG on inter-electrode distance. A group of ten healthy subjects (five males and five females) performed isometric elbow flexions of the right arm at 20, 50 and 80% of their maximal voluntary contraction (MVC). The SEMG signal was recorded using surface electrodes placed at a distance of 18 and 36 mm over the biceps brachii muscle. RMS-SEMG signals were analysed for average amplitude. At 20% MVC, no significant change was observed on the RMS value due to the spacing of the electrodes. The effect, however, was significant at 50 and 80% of MVC. Moreover, the study shows that SEMG amplitude is closely related to the level of force.

Neural networks and wavelet decomposition for classification of surface electromyography

To determine the status of a muscle, surface electromyography (SEMG) is a useful tool being non-invasive and easy to record. Clinicians are able to classify the signal visually but because of the large number of parameters of the signal, automatic classification becomes difficult. This paper reports our efforts at using wavelet transforms to process the signal before using neural networks for classification. We have found that by using specific wavelet functions and at specific levels of decomposition, the features of the signal correlating with muscle status were highlighted.