The key component in deep learning research is the availability of training data sets. With a limited number of publicly available COVID-19 chest X-ray images, the generalization and robustness of deep learning models to detect COVID-19 cases developed based on these images are questionable. We aimed to use thousands of readily available chest radiograph images with clinical findings associated with COVID-19 as a training data set, mutually exclusive from the images with confirmed COVID-19 cases, which will be used as the testing data set. We used a deep learning model based on the ResNet-101 convolutional neural network architecture, which was pretrained to recognize objects from a million of images and then retrained to detect abnormality in chest X-ray images. The performance of the model in terms of area under the receiver operating curve, sensitivity, specificity, and accuracy was 0.82, 77.3%, 71.8%, and 71.9%, respectively. The strength of this study lies in the use of labels that have a strong clinical association with COVID-19 cases and the use of mutually exclusive publicly available data for training, validation, and testing.
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.
1 August 2012 in Investigative Ophthalmology & Visual Science
Purpose.: Changes in retinal vascular parameters have been shown to be associated with systemic vascular diseases. In this study, we assessed the physiologic variations in retinal vascular measurements during the cardiac cycle. Methods.: Fundus images were taken using electrocardiogram-synchronized retinal camera at nine distinct cardiac points from 15 healthy volunteers (135 images). Analyses of retinal vessel geometric measures, including retinal vessel caliber (individual and summary), tortuosity, branching angle, length–diameter ratio (LDR), and optimality deviation, were performed using semiautomated computer software. Repeated-measures ANOVAs were used to obtain the means and to estimate the variation of each cardiac point compared with cardiac point 1. Results.: There was a significant variation of the caliber of the individual arteriolar and venular vessels. However, there was no significant variation found for vessel caliber summary, represented by the central retinal arteriolar equivalent (CRAE) and the central retinal venular equivalent (CRVE). There was also no significant variation found for tortuosity and branching angle, and LDR showed none or very little variations at different cardiac points: variations in caliber ranges between 0 and 4.1%, tortuosity 0 and 1.5%, branching angle 0 and 3.5%, and LDR 0 and 2%; all values for variations, P > 0.1; linear trend, P > 0.5; and nonlinear trend, P > 0.8. Conclusions.: This study showed that there were minimal variations in the CRAE, CRVE, tortuosity, and branching angle that are clinically used for two-dimensional measures of retinal vascular geometry during cardiac cycles. However, there was significant variation in the caliber of the individual vessels over the cardiac cycle.
This paper presents a lip reading technique to classify the discrete utterances without evaluating the acoustic signals. The reported technique analysis the video data of lip motions by computing the optical flow (OF). The statistical properties of the vertical OF component were used to form the feature vectors for training the support vector machines (SVM) classifier. The impact of the variation in speed/velocity of speaking on the performance of the system was minimized by removing the zero energy frames and normalizing the number of frames by interpolation. The resulting system is an efficient visual viseme classifier with high accuracy (95.9%), specificity (98.1%) and sensitivity (66.4%). The results of the experiments demonstrate the developed technique is insensitive to inter speaker variations.