Optical coherence tomography (OCT) and retinal fundus images are widely used for detecting retinal pathology. In particular, these images are used by deep learning methods for classification of retinal disease. The main hurdle for widespread deployment of AI-based decision making in healthcare is a lack of interpretability of the cutting-edge deep learning-based methods. Conventionally, decision making by deep learning methods is considered to be a black box. Recently, there is a focus on developing techniques for explaining the decisions taken by deep neural networks, i.e. Explainable AI (XAI) to improve their acceptability for medical applications. In this study, a framework for interpreting the decision making of a deep learning network for retinal OCT image classification is proposed. An Inception-v3 based model was trained to detect choroidal neovascularization (CNV), diabetic macular edema (DME) and drusen from a dataset of over 80,000 OCT images. We visualized and compared various interpretability methods for the three disease classes. The attributions from various approaches are compared and discussed with respect to clinical significance. Results showed a successful attribution of the specific pathological regions of the OCT that are responsible for a given condition in the absence of any pixel-level annotations.
Medical imaging datasets typically do not contain many training images and are usually not sufficient for training deep learning networks. We propose a deep residual variational auto-encoder and a generative adversarial network based approach that can generate a synthetic retinal fundus image dataset with corresponding blood vessel annotations. In terms of structural statistics comparison of real and artificial our model performed better than existing methods. The generated blood vessel structures achieved a structural similarity value of 0.74 and the artificial dataset achieved a sensitivity of 0.84 and specific city of 0.97 for the blood vessel segmentation task. The successful application of generative models for the generation of synthetic medical data will not only help to mitigate the small dataset problem but will also address the privacy concerns associated with such medical datasets.
Globally Diabetic retinopathy (DR) is one of the leading causes of blindness. But due to low patient to doctor ratio performing clinical retinal screening processes for all such patients is not always possible. In this paper a deep learning based automated diabetic retinopathy detection method is presented . Though different frameworks exist for classifying different retinal diseases with both shallow machine learning algorithms and deep learning algorithms, there is very little literature on the problem of variation of sources between training and test data. Kaggle EYEPACS data was used in this study for training the dataset and the Messidor dataset was used for testing the efficiency of the model. With proper data sampling, augmentation and pre-processing techniques it was possible to achieve state-of-the-art accuracy of classification using Messidor dataset (which had a different camera settings and resolutions of images). The model achieved significant performance with a sensitivity of almost 90% and specificity of 91. 94% with an average accuracy of 90. 4
Fundus cameras are the current clinical standard for capturing retinal images, which are used to diagnose a variety of sight-threatening conditions. Traditional fundus cameras are not easily transported, making them unsuitable for field use. In addition, traditional fundus cameras are expensive. Due to this, a variety of technologies have been developed such as the D-EYE Digital Ophthalmoscope (D-EYE Srl, Padova, Italy) which is compatible with various cellphone cameras. This paper reports on the comparison of the image quality of the Nidek RS-330 OCT Retina Scan Duo (Nidek, Tokyo, Japan) and the D-EYE paired with an iPhone 6 (Apple, Cupertino, USA). Twenty-one participants were enrolled in the study of whom 14 underwent nonmydriatic and mydriatic imaging with the D-EYE and the Nidek. Seven participants underwent nonmydriatic imaging with the D-EYE and the Nidek. The images were co-registered and cropped so that the region of interest was equal in both the D-EYE and Nidek images, as the D-EYE had a smaller field of view. Using the Nidek image as the reference, objective full-reference image quality analysis was performed. Metrics such as structural similarity index and peak signal noise ratio were obtained. It was found that the image quality of the D-EYE is limited by the attached iPhone camera, and is lower when compared to the Nidek. Quantification of the differences between the D-EYE and Nidek allows for targeted development of smartphone camera attachments that can help to bridge the current gap in image quality.
Optical Coherence tomography (OCT) images provide several indicators, e.g., the shape and the thickness of different retinal layers, which can be used for various clinical and non-clinical purposes. We propose an automated classification method to identify different ocular diseases, based on the local binary pattern features. The database consists of normal and diseased human eye SD-OCT images. We use a multiphase approach for building our classifier, including preprocessing, Meta learning, and active learning. Pre-processing is applied to the data to handle missing features from images and replace them with the mean or median of the corresponding feature. All the features are run through a Correlation-based Feature Subset Selection algorithm to detect the most informative features and omit the less informative ones. A Meta learning approach is applied to the data, in which a SVM and random forest are combined to obtain a more robust classifier. Active learning is also applied to strengthen our classifier around the decision boundary. The primary experimental results indicate that our method is able to differentiate between the normal and non-normal retina with an area under the ROC curve (AUC) of 98.6% and also to diagnose the three common retina-related diseases, i.e., Age-related Macular Degeneration, Diabetic Retinopathy, and Macular Hole, with an AUC of 100%, 95% and 83.8% respectively. These results indicate a better performance of the proposed method compared to most of the previous works in the literature.
Accurate segmentation of spectral-domain Optical Coherence Tomography (SD-OCT) images helps diagnose retinal pathologies and facilitates the study of their progression/remission. Manual segmentation is clinical-expertise dependent and highly time-consuming. Furthermore, poor image contrast due to high-reflectivity of some retinal layers and the presence of heavy speckle noise, pose severe challenges to the automated segmentation algorithms. The first step towards retinal OCT segmentation therefore, is to create a noise-free image with edge details still preserved, as achieved by image reconstruction on a wavelet-domain preceded by bilateral-filtering. In this context, the current study compares the effects of image denoising using a simple Gaussian-filter to that of wavelet-based denoising, in order to help investigators decide whether an advanced denoising technique is necessary for accurate graph-based intraretinal layer segmentation. A comparative statistical analysis conducted between the mean thicknesses of the six layers segmented by the algorithm and those reported in a previous study, reports non-significant differences for five of the layers (p > 0.05) except for one layer (p = 0.04), when denoised using Gaussian-filter. Non-significant layer thickness differences are seen between both the algorithms for all the six retinal layers (p > 0.05) when bilateral-filtering and wavelet-based denoising is implemented before boundary delineation. However, this minor improvement in accuracy is achieved at an expense of substantial increase in computation time (∼10s when run on a specific CPU) and logical complexity. Therefore, it is debatable if one should opt for advanced denoising techniques over a simple Gaussian-filter when implementing graph-based OCT segmentation algorithms.
Retinal layer shape and thickness are one of the main indicators in the diagnosis of ocular diseases. We present an active contour approach to localize intra-retinal boundaries of eight retinal layers from OCT images. The initial locations of the active contour curves are determined using a Viterbi dynamic programming method. The main energy function is a Chan-Vese active contour model without edges. A boundary term is added to the energy function using an adaptive weighting method to help curves converge to the retinal layer edges more precisely, after evolving of curves towards boundaries, in final iterations. A wavelet-based denoising method is used to remove speckle from OCT images while preserving important details and edges. The performance of the proposed method was tested on a set of healthy and diseased eye SD-OCT images. The experimental results, compared between the proposed method and the manual segmentation, which was determined by an optometrist, indicate that our method has obtained an average of 95.29%, 92.78%, 95.86%, 87.93%, 82.67%, and 90.25% respectively, for accuracy, sensitivity, specificity, precision, Jaccard Index, and Dice Similarity Coefficient over all segmented layers. These results justify the robustness of the proposed method in determining the location of different retinal layers.
Segmentation of spectral-domain Optical Coherence Tomography (SD-OCT) images facilitates visualization and
quantification of sub-retinal layers for diagnosis of retinal pathologies. However, manual segmentation is subjective,
expertise dependent, and time-consuming, which limits applicability of SD-OCT. Efforts are therefore being made to
implement active-contours, artificial intelligence, and graph-search to automatically segment retinal layers with accuracy
comparable to that of manual segmentation, to ease clinical decision-making. Although, low optical contrast, heavy
speckle noise, and pathologies pose challenges to automated segmentation. Graph-based image segmentation approach
stands out from the rest because of its ability to minimize the cost function while maximising the flow. This study has
developed and implemented a shortest-path based graph-search algorithm for automated intraretinal layer segmentation
of SD-OCT images. The algorithm estimates the minimal-weight path between two graph-nodes based on their gradients.
Boundary position indices (BPI) are computed from the transition between pixel intensities. The mean difference
between BPIs of two consecutive layers quantify individual layer thicknesses, which shows statistically insignificant
differences when compared to a previous study [for overall retina: p = 0.17, for individual layers: p > 0.05 (except one
layer: p = 0.04)]. These results substantiate the accurate delineation of seven intraretinal boundaries in SD-OCT images
by this algorithm, with a mean computation time of 0.93 seconds (64-bit Windows10, core i5, 8GB RAM). Besides
being self-reliant for denoising, the algorithm is further computationally optimized to restrict segmentation within the
user defined region-of-interest. The efficiency and reliability of this algorithm, even in noisy image conditions, makes it
clinically applicable.
Analysis of retinal fundus images is essential for physicians, optometrists and ophthalmologists in the diagnosis, care and treatment of patients. The first step of almost all forms of automated fundus analysis begins with the segmentation and subtraction of the retinal vasculature, while analysis of that same structure can aid in the diagnosis of certain retinal and cardiovascular conditions, such as diabetes or stroke. This paper investigates the use of a Convolutional Neural Network as a multi-channel classifier of retinal vessels using DRIVE, a database of fundus images. The result of the network with the application of a confidence threshold was slightly below the 2nd observer and gold standard, with an accuracy of 0.9419 and ROC of 0.9707. The output of the network with on post-processing boasted the highest sensitivity found in the literature with a score of 0.9568 and a good ROC score of 0.9689. The high sensitivity of the system makes it suitable for longitudinal morphology assessments, disease detection and other similar tasks.
The segmentation of retinal morphology has numerous applications in assessing ophthalmologic and cardiovascular disease pathologies. The early detection of many such conditions is often the most effective method for reducing patient risk. Computer aided segmentation of the vasculature has proven to be a challenge, mainly due to inconsistencies such as noise, variations in hue and brightness that can greatly reduce the quality of fundus images. Accurate fundus and/or retinal vessel maps give rise to longitudinal studies able to utilize multimodal image registration and disease/condition status measurements, as well as applications in surgery preparation and biometrics. This paper further investigates the use of a Convolutional Neural Network as a multi-channel classifier of retinal vessels using the Digital Retinal Images for Vessel Extraction database, a standardized set of fundus images used to gauge the effectiveness of classification algorithms. The CNN has a feed-forward architecture and varies from other published architectures in its combination of: max-pooling, zero-padding, ReLU layers, batch normalization, two dense layers and finally a Softmax activation function. Notably, the use of Adam to optimize training the CNN on retinal fundus images has not been found in prior review. This work builds on prior work of the authors, exploring the use of Gabor filters to boost the accuracy of the system to 0.9478 during post processing. The mean of a series of Gabor filters with varying frequencies and sigma values are applied to the output of the network and used to determine whether a pixel represents a vessel or non-vessel.
Visual computations such as depth-from-stereo are highly dependent on edges and textures for the process of image correspondence. IR images typically lack the necessary detail for producing dense depth maps, however, sparse maps may be adequate for autonomous obstacle avoidance. We have constructed an IR stereo head for eventual UGV and UAV night time navigation. In order to calibrate the unit, we have constructed a thermal calibration checkerboard. We show that standard stereo camera calibration based on a checkerboard developed for calibrating visible spectrum cameras can also be used for calibrating an IR stereo pair, with of course hot/cold squares used as opposed to black/white squares. Once calibrated, the intrinsic and extrinsic parameters for each camera provide the absolute depth value if a left-right correspondence can be established. Given the general texture-less characteristic of IR imagery, selecting key salient features that are left-right stable and tractable is key for producing a sparse depth map. IR imagery, like visible and range maps is highly spatially correlated and a dense map can be obtained from a sparse map via propagation. Preliminary results from salient IR feature detection are investigated as well.
KEYWORDS: Data fusion, Neural networks, Testing and analysis, Databases, Detection and tracking algorithms, Imaging systems, Image fusion, Feature extraction, Digital watermarking, Control systems
Static signature verification is a well researched problem that has not been completely solved to this date. To improve on current verification performance this research uses a pooling method which fuses together decisions of selected verification algorithms. To enhance this performance further, the decision from this method is fused with the decision of a neural network classifier. This neural network classifier offers a new approach to signature verification, since it is based on recognition techniques. The advantage of this classifier is that it incorporates different information into its decision and therefore allows the fused decision to be based on more diverse information. In contrast to other methods, this classifier requires only genuine signature samples to be trained. Experimental results show that the fusion of verification algorithms can produce better performance than any of the used methods individually.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.