Deep learning detection of diabetic retinopathy in Scotland's diabetic eye screening programme.

BACKGROUND/AIMS: Support vector machine-based automated grading (known as iGradingM) has been shown to be safe, cost-effective and robust in the diabetic retinopathy (DR) screening (DES) programme in Scotland. It triages screening episodes as gradable with no DR versus manual grading required. The study aim was to develop a deep learning-based autograder using images and gradings from DES and to compare its performance with that of iGradingM. METHODS: Retinal images, quality assurance (QA) data and routine DR grades were obtained from national datasets in 179 944 patients for years 2006-2016. QA grades were available for 744 images. We developed a deep learning-based algorithm to detect whether either eye contained ungradable images or any DR. The sensitivity and specificity were evaluated against consensus QA grades and routine grades. RESULTS: Images used in QA which were ungradable or with DR were detected by deep learning with better specificity compared with manual graders (p<0.001) and with iGradingM (p<0.001) at the same sensitivities. Any DR according to the DES final grade was detected with 89.19% (270 392/303 154) sensitivity and 77.41% (500 945/647 158) specificity. Observable disease and referable disease were detected with sensitivities of 96.58% (16 613/17 201) and 98.48% (22 600/22 948), respectively. Overall, 43.84% of screening episodes would require manual grading. CONCLUSION: A deep learning-based system for DR grading was evaluated in QA data and images from 11 years in 50% of people attending a national DR screening programme. The system could reduce the manual grading workload at the same sensitivity compared with the current automated grading system.

Automated clarity assessment of retinal images using regionally based structural and statistical measures.

An automated image analysis system for application in mass medical screening must assess the clarity of the images before analysing their content. This is the case in grading for diabetic retinopathy screening where the failure to assess clarity could result in retinal images of people with retinopathy being erroneously classed as normal. This paper compares methods of clarity assessment based on the degradation of visible structures and based on the deviation of image properties outside expected norms caused by clarity loss. Vessel visibility measures and statistical measures were determined at locations in the image which have high saliency and these were used to obtain an image clarity assessment using supervised classification. The usefulness of the measures as indicators of image clarity was assessed. Tests were performed on 987 disc-centred and macula-centred retinal photographs (347 with inadequate clarity) obtained from the English National Screening Programme. Images with inadequate clarity were detected with 92.6% sensitivity at 90% specificity. In a set of 2000 macula-centred images (200 with inadequate clarity) from the Scottish Screening Programme, inadequate clarity was detected with 96.7% sensitivity at 90% specificity. This study has shown that structural and statistical measures are equally useful for retinal image clarity assessment.

The evidence for automated grading in diabetic retinopathy screening.

Systematic screening for diabetic retinopathy using retinal photography has been shown to reduce the incidence of blindness among people with diabetes. The implementation of diabetic retinopathy screening programmes faces several challenges. Consequently, methods for improving the efficiency of screening are being sought, one of which is the automation of image grading involving detection of images with either disease or of inadequate quality using computer software. This review aims to bring together the available evidence that is suitable for making a judgement about whether automated grading systems could be used effectively in diabetic retinopathy screening. To do this, it focuses on studies made by the few centres who have presented results tests of automated grading software on large sets of patients or screening episodes. It also considers economic model analyses and papers describing the effectiveness of manual grading in order that the effect of replacing stages of manual grading by automated grading can be judged. In conclusion, the review shows that there is sufficient evidence to suggest that automated grading, operating as a disease / no disease grader, is safe and could reduce the workload of manual grading in diabetic retinopathy screening.

Detection of new vessels on the optic disc using retinal photographs.

Proliferative diabetic retinopathy is a rare condition likely to lead to severe visual impairment. It is characterized by the development of abnormal new retinal vessels. We describe a method for automatically detecting new vessels on the optic disc using retinal photography. Vessel-like candidate segments are first detected using a method based on watershed lines and ridge strength measurement. Fifteen feature parameters, associated with shape, position, orientation, brightness, contrast and line density are calculated for each candidate segment. Based on these features, each segment is categorized as normal or abnormal using a support vector machine (SVM) classifier. The system was trained and tested by cross-validation using 38 images with new vessels and 71 normal images from two diabetic retinal screening centers and one hospital eye clinic. The discrimination performance of the fifteen features was tested against a clinical reference standard. Fourteen features were found to be effective and used in the final test. The area under the receiver operator characteristic curve was 0.911 for detecting images with new vessels on the disc. This accuracy may be sufficient for it to play a useful clinical role in an automated retinopathy analysis system.

Automated grading for diabetic retinopathy: a large-scale audit using arbitration by clinical experts.

BACKGROUND/AIMS: Automated grading software has the potential to reduce the manual grading workload within diabetic retinopathy screening programmes. This audit was undertaken at the request of Scotland's National Diabetic Retinopathy Screening Collaborative to assess whether the introduction of automated grading software into the national screening programme would be safe, robust and effective. METHODS: Automated grading, performed by software for image quality assessment and for microaneurysm/dot haemorrhage detection, was carried out on 78,601 images, obtained from 33,535 consecutive patients, which had been manually graded at one of two regional diabetic retinopathy screening programmes. Cases where the automated grading software assessment indicated gradable images with no disease but the screening programme indicated ungradable images or disease more severe than mild retinopathy were arbitrated by seven senior ophthalmologists. RESULTS: 100% (180/180) of patients with proliferative retinopathy, 100% (324/324) with referable background retinopathy, 100% (193/193) with observable background retinopathy, 97.3% (1099/1130) with referable maculopathy, 99.2% (384/387) with observable maculopathy and 99.8% (1824/1827) with ungradable images were detected by the software. CONCLUSION: The automated grading software operated to previously published results when applied to a large, unselected population attending two regional screening programmes. Manual grading workload reduction would be 36.3%.

The role of haemorrhage and exudate detection in automated grading of diabetic retinopathy.

BACKGROUND/AIMS: Automated grading has the potential to improve the efficiency of diabetic retinopathy screening services. While disease/no disease grading can be performed using only microaneurysm detection and image-quality assessment, automated recognition of other types of lesions may be advantageous. This study investigated whether inclusion of automated recognition of exudates and haemorrhages improves the detection of observable/referable diabetic retinopathy. METHODS: Images from 1253 patients with observable/referable retinopathy and 6333 patients with non-referable retinopathy were obtained from three grading centres. All images were reference-graded, and automated disease/no disease assessments were made based on microaneurysm detection and combined microaneurysm, exudate and haemorrhage detection. RESULTS: Introduction of algorithms for exudates and haemorrhages resulted in a statistically significant increase in the sensitivity for detection of observable/referable retinopathy from 94.9% (95% CI 93.5 to 96.0) to 96.6% (95.4 to 97.4) without affecting manual grading workload. CONCLUSION: Automated detection of exudates and haemorrhages improved the detection of observable/referable retinopathy.

Automated detection of exudates for diabetic retinopathy screening.

Automated image analysis is being widely sought to reduce the workload required for grading images resulting from diabetic retinopathy screening programmes. The recognition of exudates in retinal images is an important goal for automated analysis since these are one of the indicators that the disease has progressed to a stage requiring referral to an ophthalmologist. Candidate exudates were detected using a multi-scale morphological process. Based on local properties, the likelihoods of a candidate being a member of classes exudate, drusen or background were determined. This leads to a likelihood of the image containing exudates which can be thresholded to create a binary decision. Compared to a clinical reference standard, images containing exudates were detected with sensitivity 95.0% and specificity 84.6% in a test set of 13,219 images of which 300 contained exudates. Depending on requirements, this method could form part of an automated system to detect images showing either any diabetic retinopathy or referable diabetic retinopathy.

Automatic detection of retinal anatomy to assist diabetic retinopathy screening.

Screening programmes for diabetic retinopathy are being introduced in the United Kingdom and elsewhere. These require large numbers of retinal images to be manually graded for the presence of disease. Automation of image grading would have a number of benefits. However, an important prerequisite for automation is the accurate location of the main anatomical features in the image, notably the optic disc and the fovea. The locations of these features are necessary so that lesion significance, image field of view and image clarity can be assessed. This paper describes methods for the robust location of the optic disc and fovea. The elliptical form of the major retinal blood vessels is used to obtain approximate locations, which are refined based on the circular edge of the optic disc and the local darkening at the fovea. The methods have been tested on 1056 sequential images from a retinal screening programme. Positional accuracy was better than 0.5 of a disc diameter in 98.4% of cases for optic disc location, and in 96.5% of cases for fovea location. The methods are sufficiently accurate to form an important and effective component of an automated image grading system for diabetic retinopathy screening.

Automated microaneurysm detection using local contrast normalization and local vessel detection.

Screening programs using retinal photography for the detection of diabetic eye disease are being introduced in the UK and elsewhere. Automatic grading of the images is being considered by health boards so that the human grading task is reduced. Microaneurysms (MAs) are the earliest sign of this disease and so are very important for classifying whether images show signs of retinopathy. This paper describes automatic methods for MA detection and shows how image contrast normalization can improve the ability to distinguish between MAs and other dots that occur on the retina. Various methods for contrast normalization are compared. Best results were obtained with a method that uses the watershed transform to derive a region that contains no vessels or other lesions. Dots within vessels are handled successfully using a local vessel detection technique. Results are presented for detection of individual MAs and for detection of images containing MAs. Images containing MAs are detected with sensitivity 85.4% and specificity 83.1%.

Automated assessment of diabetic retinal image quality based on clarity and field definition.

PURPOSE: To evaluate the performance of an automated retinal image quality assessment system for use in automated diabetic retinopathy grading. METHODS: Algorithmic methods have been developed for assessing the quality of 45 degrees single field retinal images for use in diabetic retinopathy screening. For this purpose, image quality was defined by two aspects: image clarity and field definition. An image with adequate clarity was defined as one that shows sufficient detail for automated retinopathy grading. The visibility of the macular vessels was used as an indicator of image clarity, since these vessels are known to be narrow and become less visible with any image degradation. An image with adequate field definition was defined as one that shows the desired field of view for retinopathy grading, including the full 45 degrees field of view, the optic disc, and at least two optic disc diameters of visible retina around the fovea. From 489 patients attending a diabetic retinopathy screening program, 1039 retinal images were obtained. The images were graded by a clinician for image clarity and field definition, with a comprehensive image-quality grading scheme. RESULTS: The sensitivity and specificity were, respectively, 100% and 90.9% for inadequate clarity detection, 95.3% and 96.4% for inadequate field definition detection, and 99.1% and 89.4% for inadequate overall quality detection. CONCLUSIONS: The automated system performs with sufficient accuracy to form part of an automated diabetic retinopathy grading system.

SUPEROVULATION AND SUPERPREGNANCY IN THE GOLDEN HAMSTER.

Pregnant mare serum gonadotropin (PMSG) treatment given in the morning or afternoon on any day of the four-day estrous cycle and human chorionic gonado tropin (HCG) given two days later successfully induced superovulation in the golden hamster. The minimum interval between PMSG and HCG necessary to obtain consistent superovulation was approximately 44 hr. The lowest ovulation rate was obtained following PMSG treatment on the afternoon of day 4 despite the fact that this time coincides with the maximum endogenous FSH level, necessary for the maturation of the next crop of follicles destined to ovulate. Thirty-eight to one-hundred percent of superovulated females in four different treatment groups became superpregnant after natural mating. Some treated females exhibited two consecutive nights of estrus with ovulation apparently occurring during the second night. Superpregnant females delivered "super" size litters, up to 27 live-born pups. The ultimate litter size appeared to be established after day 3 and prior to day 8 of superpregnancy. A one-day extension of the normal 16-day gestation period was observed in 31% of superpregnancies. Unilateral pregnancies were observed at autopsy in 44% of treated females which received the high dose of PMSG (30 IU). The progeny of superovulated females reproduced normally at maturity. The results indicate that ova from superovulated female hamsters are capable of full normal development.