Screening and prediction of diabetes using retinal images by artificial intelligence
Screening and prediction of diabetes using retinal images by artificial intelligence
Ahmad Nejati Shahidain,1Akram Hoseinzadeh,2Mohammad Mahdi Khalizadeh,3,*
1. Department of Biomedical engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran 2. Department of Immunology, Faculty of Medicine, Mashhad University of Medical Sciences, Mashhad, Iran 3. Department of Biomedical engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran
Introduction: Diabetes is a metabolic disorder that leads to complications including cardiovascular renal and eye disease. Diabetes is a significant and costly heath problem in the world and is growing in incidence at almost epidemic levels. New and innovative ways of identification diagnosis treatment and follow up are needed to manage this growing problem. Using the retina and its blood vessel characteristics can provide a window into several disease processes. morphological characteristic the change in the venous or arterial vessel diameter especially in proximity to the optic disc allows the application of image analysis and automated classification in risk assessment of diabetes disease. Detection of width changes in blood vessels of the retina may be indicative of eye or systemic disease.
Methods: In order to detect vessel width change vessels must first be identified in available digital images. Extraction algorithms generally use exploratory techniques. They are faster computationally and usually determining useful morphometric information as part of the discovery process. Since the vessel boundaries are part of the discovery process these algorithms generally contain information such as vessel widths center points, and local orientation. at width change detection would need to follow three basic steps. The first step is to find the vessels in each image with the boundaries being identified to subpixel accuracy. Second is the step of transforming all vessels into the same coordinate system and identifying corresponding vessel pieces. Last is the ability to measure to subpixel accuracy the vessel widths and to identify changes in width over time. By using the DBICP registration algorithm, it is possible to determine transformations that can be used to accurately align both images as well as results generated from these images. Once final blood vessel boundary locations are determined, it is then possible to transform images into common coordinate systems through the process of registration. Once registered widths can be compared utilizing the described hypothesis test framework attributing any change of 5 percent or less to normal vessel changes caused by the cardiac rhythm. For an experimental study a fixed number of subjects with known disease and a fixed number of no diseased subjects are selected and the diagnostic test is applied to both categories. This is a low-cost design useful for early stages of development of a new test.
Results: A preliminary validation of this software showed a sensitivity and specificity of 80 percent for the detection of normality based on precise detection of individual lesions. Decisions that depend on the detection of lesion patterns such as clinically significant macular edema showed a sensitivity and specificity of more than 95 percent. Validation by two expert graders suggested a sensitivity and specificity of below 90 percent for any lesion and of more than 95 percent for predicting overall retinopathy grade. The features of this system include high operating speed, performing a large number of samples at the same time the ability to send images to remote locations etc.
Conclusion: Such of the clinical investigation related to retinal vascular geometry has focused on hypertension and cardiovascular disease but less attention so far has been given to other systemic diseases such as diabetes. Early unpublished findings in diabetic patients suggest that retinal vascular geometry may also yield clinically useful information here. Such developments offer the prospect of a highly automated screening tool suitable for centralized analysis of retinal photographs captured locally using existing digital fundus cameras for example during routine examinations by optometrists. Such an arrangement might offer a highly cost effective opportunity to screen large populations for risk factors in a range of systemic diseases.