Gilberto Zamora

Automated image quality evaluation of retinal fundus photographs in diabetic retinopathy screening

This paper presents a system that can automatically determine whether the quality of a retinal image is sufficient for computer-based diabetic retinopathy (DR) screening. The system integrates global histogram features, textural features, and vessel density, as well as a local non-reference perceptual sharpness metric. A partial least square (PLS) classifier is trained to distinguish low quality images from normal quality images. The system was evaluated on a large, representative set of 1884 non-mydriatic retinal images from 412 subjects. An area under the ROC curve of 96% was achieved.

Lossless coding of color images using color space transformations

The key issue in the development of lossless compression techniques has been the manipulation of the image model from a probabilistic point of view in order to reduce the bit rates as represented by the entropy of such a model. In this respect, arithmetic coding using adaptive models shows excellent performance and is easy to implement. On the other hand, many studies have been dedicated to find suitable ways to represent a color image based on the human visual response to color changes in magnitude, frequency and direction. Many of these studies have resulted in linear and nonlinear transformations of the well-known RGB and tristimulus color spaces that describe a color image in different planes, each of those being a function of the illuminance and chrominance of each pixel. Application of lossless coding techniques to these transformed color planes, such as L*a*b*, has resulted in improved performances compared to those in RGB planes.

A robust registration technique for multi-sensor images

This paper describes an improved technique to register multi-sensor images by segmenting the images by adaptive clustering prior to performing preprocessing and cepstrum operation to determine the translational displacement. The difficulty in registering multi sensor images lies in the fact that the images of the same scene acquired by different sensors often appear different in detailed structures. Therefore the common features existing in such images need to be identified by suitable preprocessing operations for the success of the cepstral registration technique. Experimental results demonstrate the feasibility of successful cepstral registration of SAR and electro-optic images of the same scene despite apparent noticeable differences in some embedded structures thus providing a potential powerful tool for automated registration.

Lossless compression of segmented biomedical images with an adaptive arithmetic coding model

Any image can be segmented into regions containing information without the background that can be added at the reconstruction stage thus providing a lossless compression of approximately 3:1 in most cases. When this segmentation process is combined with an adaptive arithmetic coding model, an extremely efficient and powerful lossless coding model yielding compression ratios up to 9:1 can be developed. Such models can be quite useful for archiving wide varieties of biomedical images losslessly.

Segmentation by color space transformation prior to lifting and integer wavelet transformation for efficient lossless coding and transmission

Efficient lossless coding and transmission of large color images through bandlimited channels are of growing interest in many applications. Classically lossless coding is performed in the spatial domain by entropy coding of the decorrelated pixels. However, such coding yields only a maximum of 3:1 compression. Recent developments in integer to integer wavelet transform allow multiresolution representation and progressive transmission of images through bandlimited channels starting from the entropy coded lowest resolution image. However, such schemes are able to reduce the bit rate to a limited extent. We present here an efficient automated technique using a simple color space transformation from RGB to HSI that segments and retrieves the color object contour in a fast manner. When such a segmented image is subjected to the integer to integer wavelet transform followed by an adaptive arithmetic coding model, lossless compression up to 20:1 has been achieved for some color images. Due to the substantial decrease in size of the input image to be processed, the execution time of the entire algorithm is reduced drastically. Although the color transformation from RGB to L/sup */a/sup */b/sup */ results in better visual appearance as well as more compressibility, this nonlinear transformation is more computationally intensive. Therefore for progressive transmission a simple color transformation scheme is preferable.

Comparison of low-cost handheld retinal camera and traditional table top retinal camera in the detection of retinal features indicating a risk of cardiovascular disease

Retinal abnormalities associated with hypertensive retinopathy are useful in assessing the risk of cardiovascular disease, heart failure, and stroke. Assessing these risks as part of primary care can lead to a decrease in the incidence of cardiovascular disease-related deaths. Primary care is a resource limited setting where low cost retinal cameras may bring needed help without compromising care. We compared a low-cost handheld retinal camera to a traditional table top retinal camera on their optical characteristics and performance to detect hypertensive retinopathy. A retrospective dataset of N=40 subjects (28 with hypertensive retinopathy, 12 controls) was used from a clinical study conducted at a primary care clinic in Texas. Non-mydriatic retinal fundus images were acquired using a Pictor Plus hand held camera (Volk Optical Inc.) and a Canon CR1-Mark II tabletop camera (Canon USA) during the same encounter. The images from each camera were graded by a licensed optometrist according to the universally accepted Keith-Wagener-Barker Hypertensive Retinopathy Classification System, three weeks apart to minimize memory bias. The sensitivity of the hand-held camera to detect any level of hypertensive retinopathy was 86% compared to the Canon. Insufficient photographer’s skills produced 70% of the false negative cases. The other 30% were due to the handheld camera’s insufficient spatial resolution to resolve the vascular changes such as minor A/V nicking and copper wiring, but these were associated with non-referable disease. Physician evaluation of the performance of the handheld camera indicates it is sufficient to provide high risk patients with adequate follow up and management.

Low cost thermal camera for use in preclinical detection of diabetic peripheral neuropathy in primary care setting

Diabetic peripheral neuropathy (DPN) accounts for around 73,000 lower-limb amputations annually in the US on patients with diabetes. Early detection of DPN is critical. Current clinical methods for diagnosing DPN are subjective and effective only at later stages. Until recently, thermal cameras used for medical imaging have been expensive and hence prohibitive to be installed in primary care setting. The objective of this study is to compare results from a low-cost thermal camera with a high-end thermal camera used in screening for DPN. Thermal imaging has demonstrated changes in microvascular function that correlates with nerve function affected by DPN. The limitations for using low-cost cameras for DPN imaging are: less resolution (active pixels), frame rate, thermal sensitivity etc. We integrated two FLIR Lepton (80x60 active pixels, 50° HFOV, thermal sensitivity < 50mK) as one unit. Right and left cameras record the videos of right and left foot respectively. A compactible embedded system (raspberry pi3 model Bv1.2) is used to configure the sensors, capture and stream the video via ethernet. The resulting video has 160x120 active pixels (8 frames/second). We compared the temperature measurement of feet obtained using low-cost camera against the gold standard highend FLIR SC305. Twelve subjects (aged 35-76) were recruited. Difference in the temperature measurements between cameras was calculated for each subject and the results show that the difference between the temperature measurements of two cameras (mean difference=0.4, p-value=0.2) is not statistically significant. We conclude that the low-cost thermal camera system shows potential for use in detecting early-signs of DPN in under-served and rural clinics.

Transfer learning for diabetic retinopathy

Diabetic Retinopathy (DR)1, 2 is a leading cause of blindness worldwide and is estimated to threaten the vision of nearly 200 million by 2030.3 To work with the ever-increasing population, the use of image processing algorithms to screen for those at risk has been on the rise. Research-oriented solutions have proven effective in classifying images with or without DR, but often fail to address the true need of the clinic - referring only those who need to be seen by a specialist, and reading every single case. In this work, we leverage an array of image pre-preprocessing techniques, as well as Transfer Learning to re-purpose an existing deep network for our tasks in DR. We train, test, and validate our system on 979 clinical cases, achieving a 95% Area Under the Curve (AUC) for referring Severe DR with an equal error Sensitivity and Specificity of 90%. Our system does not reject any images based on their quality, and is agnostic in terms of eye side and field. These results show that general purpose classifiers can, with the right type of input, have a major impact in clinical environments or for teams lacking access to large volumes of data or high-throughput supercomputers.

Portable, low-priced retinal imager for eye disease screening

The objective of this project was to develop and demonstrate a portable, low-priced, easy to use non-mydriatic retinal camera for eye disease screening in underserved urban and rural locations. Existing portable retinal imagers do not meet the requirements of a low-cost camera with sufficient technical capabilities (field of view, image quality, portability, battery power, and ease-of-use) to be distributed widely to low volume clinics, such as the offices of single primary care physicians serving rural communities or other economically stressed healthcare facilities. Our approach for Smart i-Rx is based primarily on a significant departure from current generations of desktop and hand-held commercial retinal cameras as well as those under development. Our techniques include: 1) Exclusive use of off-the-shelf components; 2) Integration of retinal imaging device into low-cost, high utility camera mount and chin rest; 3) Unique optical and illumination designed for small form factor; and 4) Exploitation of autofocus technology built into present digital SLR recreational cameras; and 5) Integration of a polarization technique to avoid the corneal reflex. In a prospective study, 41 out of 44 diabetics were imaged successfully. No imaging was attempted on three of the subjects due to noticeably small pupils (less than 2mm). The images were of sufficient quality to detect abnormalities related to diabetic retinopathy, such as microaneurysms and exudates. These images were compared with ones taken non-mydriatically with a Canon CR-1 Mark II camera. No cases identified as having DR by expert retinal graders were missed in the Smart i-Rx images.

Analysis of the feasibility of using active shape models for segmentation of gray-scale images

Active Shape Models (ASM) have been used extensively to segment images where the objects of interest show little to moderate shape variability across a training set. It is well known that the efficacy of this technique relies heavily on the quality of the training set and the initialization of the mean shape on the target image. However, little has been said about the validity of the assumptions under which the two core components of ASM, i.e. the shape model and the gray level model, are built. We explore these assumptions and test their validity with respect to both shape and gray level models. In this study, we use different training sets of real and synthetic gray scale images and investigate the reasons for their success or failure in the context of shape and gray level modeling. We show that the shape model performance is not affected by small changes in the distribution of the shapes. Furthermore, we show that a reason for segmentation failure is the lack of features in the mean profiles of gray level values that causes localization errors even under ideal conditions.