Thomas P. Karnowski et al.

Fig. 14.4. Sample encrypted patient report with electronic signature and certificate validation.

estimate that the automatic diagnosis can be assigned within 30 seconds. The average period (and variance) for each processing step is shown in Table 14.1. We estimate that the average response time from image acquisition to the generation of a diagnostic and management report for the end user is less than two minutes. The network structure provides sufficiently high throughput for image analysis and diagnosis of diabetic eye disease.

14.2. Image Analysis

The use of computer-based image analysis in ophthalmology has been a subject of research for some time. The use of telemedicine employing

428

Automating the Diagnosis, Stratification, and Management of DR

Table 14.1. Average response time from image acquisition to generation of a diagnostic report in a fully automated system.

human experts to diagnose DR from digital images is beginning to grow. Established retinal reading centers such as the Joslin Vision Network (Boston, MA) have shown that digital photography is an excellent tool for identifying DR when performed by experienced and certified readers.24,25 In addition, a web-based telemedicine system for DR with open-source software has been demonstrated and other concepts have been explored.26,27

There has also been considerable research in the development of automated screening for DR. An excellent overview of this area is provided in Ref. [10], which summarizes image segmentation as a two-step process. The first step is summarized as identifying the regular or expected physiological features of the retina. The second step is concerned with identifying the pathology of the retina through the detection of abnormal phenomena (lesions). The final step uses an intelligent system to analyze and diagnose the segmentation results. Recent reports on this subject include Ref. [28], which describes information fusion and the use of a computer-aided diagnosis (CAD) system for the detection of normal and abnormal fundus images.

In our system, image processing and analysis falls into a sequence of steps, which are detailed as follows and shown in Fig. 14.5. First, quality estimation is performed on the image at the clinical site to ensure adequate image quality. Vessel segmentation is conducted in conjunction with this operation and is saved for further analysis. The image and examination

429

Thomas P. Karnowski et al.

Fig. 14.5. Information flow in the telemedical network. The left path is manually driven and the right path is processed automatically.

information is saved in a database as described earlier for use in physician diagnosis, typically using a network-interfaced PC or smart phone. In a parallel track, the processing flow is automated for computer diagnosis of the retinal images. The vessel segmentation is used to estimate the location of the ON in each image. A complementary analysis is performed using an alternate ON detection method, and the results are compared to present a confidence level on the ON detection. The location of the macula is estimated using a nonlinear parabolic fit to the vessel tree and statistics on retina physiology. Lesions are detected next: our current work focuses on microaneurysms and exudates, as they are the main indicators of DR. The final estimated lesion population is then measured using numerical descriptors, which are then projected to a custom, semantic-driven index space. This “directed index” space allows the rapid grouping of similar images and, thus, permits our system to ultimately search through hundreds of thousands of examples using CBIR. The retrieved candidate images are then brute-force searched using distance metrics measured on the individual feature vectors. Confidence measurements (which will be described) are attached to the retrieval based on the population of the image database. In a

430

Automating the Diagnosis, Stratification, and Management of DR

complete system, computer diagnosis can replace an ophthalmologist, when some level of physician oversight is built-in for statistical sampling, or when confidence levels do not exceed an acceptable level. In our research system, the computer-generated reports are analyzed and compared with the actual physician diagnosis for statistical comparison and system improvement.

14.2.1. Vascular Tree Segmentation

Segmentation refers to the process of labeling an image to distinguish between the background and foreground of objects of interest. In vascular tree segmentation, our goal is to identify the vessels in the retinal image as distinct from the background or nonvessel pixels. There has been much research on this area in the literature dating back several years with newer work published fairly frequently; see for example Refs. [29–33]. Much of the interest in vessel segmentation stems from the argument that some ocular diseases, such as retinal vein occlusions and hypertensive retinopathy can be diagnosed through the analysis of the vascular tree (Fig. 14.6). The vessels can also serve to provide landmarks for the registration of multiple images of the same retina from different fields of view or from different time instances. Finally, they may also be used to locate key physiological elements in the retina, such as the ON and fovea. A valuable comparison of retina-vessel segmentation methods was performed in Ref. [34] in which medical experts hand-segmented a publicly available database. Their results were then compared to automatic segmentation methods. The comparison method was the agreement between pixel classification (as part of the vessel or the background) for one manual segmentation method and the automated

Fig. 14.6. Example retinal images showing an abnormal superior vascular tree of the left eye from a hemi-retinal vein occlusion.

431