Universal adversarial attacks on deep neural networks for medical image classification.

BACKGROUND: Deep neural networks (DNNs) are widely investigated in medical image classification to achieve automated support for clinical diagnosis. It is necessary to evaluate the robustness of medical DNN tasks against adversarial attacks, as high-stake decision-making will be made based on the diagnosis. Several previous studies have considered simple adversarial attacks. However, the vulnerability of DNNs to more realistic and higher risk attacks, such as universal adversarial perturbation (UAP), which is a single perturbation that can induce DNN failure in most classification tasks has not been evaluated yet. METHODS: We focus on three representative DNN-based medical image classification tasks (i.e., skin cancer, referable diabetic retinopathy, and pneumonia classifications) and investigate their vulnerability to the seven model architectures of UAPs. RESULTS: We demonstrate that DNNs are vulnerable to both nontargeted UAPs, which cause a task failure resulting in an input being assigned an incorrect class, and to targeted UAPs, which cause the DNN to classify an input into a specific class. The almost imperceptible UAPs achieved > 80% success rates for nontargeted and targeted attacks. The vulnerability to UAPs depended very little on the model architecture. Moreover, we discovered that adversarial retraining, which is known to be an effective method for adversarial defenses, increased DNNs' robustness against UAPs in only very few cases. CONCLUSION: Unlike previous assumptions, the results indicate that DNN-based clinical diagnosis is easier to deceive because of adversarial attacks. Adversaries can cause failed diagnoses at lower costs (e.g., without consideration of data distribution); moreover, they can affect the diagnosis. The effects of adversarial defenses may not be limited. Our findings emphasize that more careful consideration is required in developing DNNs for medical imaging and their practical applications.

Grading Diabetic Retinopathy from Stereoscopic Color Fundus Photographs - An Extension of the Modified Airlie House Classification: ETDRS Report Number 10.

The modified Airlie House classification of diabetic retinopathy has been extended for use in the Early Treatment Diabetic Retinopathy Study (ETDRS). The revised classification provides additional steps in the grading scale for some characteristics, separates other characteristics previously combined, expands the section on macular edema, and adds several characteristics not previously graded. The classification is described and illustrated and its reproducibility between graders is assessed by calculating percentages of agreement and kappa statistics for duplicate gradings of baseline color non- simultaneous stereoscopic fundus photographs. For retinal hemorrhages and/ or microaneurysms, hard exudates, new vessels, fibrous proliferations, and macular edema, agreement was substantial (weighted kappa, 0.61 to 0.80). For soft exudates, intraretinal microvascular abnormalities, and venous beading, agreement was moderate (weighted kappa, 0.41 to 0.60). A double grading system, with adjudication of disagreements of two or more steps between duplicate gradings, led to some improvement in reproducibility for most characteristics.

SURGICAL OUTCOMES OF MYOPIC TRACTION MACULOPATHY ACCORDING TO THE INTERNATIONAL PHOTOGRAPHIC CLASSIFICATION FOR MYOPIC MACULOPATHY.

PURPOSE: To compare treatment results of myopic traction maculopathy according to the international photographic classification for myopic maculopathy. METHODS: This was a retrospective, single-surgeon-based, observational case series of 35 consecutive eyes that underwent vitrectomy for myopic traction maculopathy. Eyes were classified into nonpathologic myopia (PM) (n = 15) and PM (n = 20) groups. Main outcome measures constituted best-corrected visual acuity (BCVA) and anatomical change. RESULTS: The mean follow-up was 32.03 ± 6.85 months. Axial length correlated with myopic maculopathy category (rho = 0.6836, P < 0.001). In the total group, BCVA improved from 20/61 to 20/36 (P = 0.001). In the subgroup, BCVA improved from 20/41 to 20/22 in the non-PM group (P = 0.002), whereas from 20/82 to 20/52 in the PM group (P = 0.048). Postoperative BCVA of the PM group was inferior to that of the non-PM group (P = 0.002) and the PM group was more likely to have postoperative BCVA <20/30 (odds ratio, 17.3; 95% CI, 2.6-325.0; P = 0.012). Two cases of macular hole retinal detachment occurred after surgery in the PM group. CONCLUSION: Because there are limited benefits of vitrectomy in myopic traction maculopathy accompanied by PM, careful consideration would be necessary when determining surgery. Optical coherence tomography should not be used alone in determining vitrectomy because myopic traction maculopathy can also have PM defined mainly by fundus photographs.

Validity of a new comprehensive pterygia grading scale for use in clinical research and clinical trial.

PURPOSE: To determine the reliability and agreement of a new comprehensive pterygium grading scale for use in clinical research and clinical trials. METHODS: Thirty eyes with pterygia were enrolled in this study. Primary gaze position and lateral gaze position images were taken of each eye with a modified single-lens reflex camera system. Our grading scale includes five parameters: two hyperemia parameters of pterygia on two different gaze position images and three size parameters, quantifying length, width, and area of the cornea encroachment of pterygium, using ImageJ software. All images were graded on the five parameters by two masked, certified reading center graders. Two graders independently graded all the images to determine inter-grader reliability. One grader regraded the images after 3 days to determine intra-grader reliability. Intraclass correlation coefficient (ICC) and inter-rater agreement statistic (κ) calculations were performed. RESULTS: The intra-grader reliability for hyperemia grading was high on both primary and lateral gazing positions (κ value is 0.93 and 0.96). The inter-grader reliability for hyperemia grading was also good (κ value is 0.85 and 0.87). The mean value of width, length, and area of the cornea encroachment of pterygium was 4.31 ± 2.04 mm, 2.08 ± 1.43 mm, and 7.84 ± 7.62 mm2, respectively. The intra-grader agreement on width, length, and area were excellent, with ICCs of 0.98 (95% CI 0.96-0.99), 0.99 (95% CI 0.98-1.0), and 0.97 (95% CI 0.94-0.99), respectively. The inter-grader agreement on width, length, and area were also excellent, with ICCs of 0.96 (95% CI 0.90-0.98), 0.99 (95% CI 0.98-0.99), and 0.99 (95% CI 0.97-0.99), respectively. CONCLUSIONS: There was excellent intra- and inter-observer reproducibility with the new comprehensive grading scale. This scale could lead to the development of standardized grading assessments and quantification of pterygia that would be valid in clinical research and clinical trials.

BMC ecology image competition 2017: the winning images.

For the fifth year, BMC Ecology is proud to present the winning images from our annual image competition. The 2017 edition received entries by talented shutterbug-ecologists from across the world, showcasing research that is increasing our understanding of ecosystems worldwide and the beauty and diversity of life on our planet. In this editorial we showcase the winning images, as chosen by our Editorial Board and guest judge Chris Darimont, as well as our selection of highly commended images. Enjoy!

Identification of Diabetic Retinopathy and Ungradable Image Rate with Ultrawide Field Imaging in a National Teleophthalmology Program.

PURPOSE: To compare diabetic retinopathy (DR) identification and ungradable image rates between nonmydriatic ultrawide field (UWF) imaging and nonmydriatic multifield fundus photography (NMFP) in a large multistate population-based DR teleophthalmology program. DESIGN: Multiple-site, nonrandomized, consecutive, cross-sectional, retrospective, uncontrolled imaging device evaluation. PARTICIPANTS: Thirty-five thousand fifty-two eyes (17 526 patients) imaged using NMFP and 16 218 eyes (8109 patients) imaged using UWF imaging. METHODS: All patients undergoing Joslin Vision Network (JVN) imaging with either NMFP or UWF imaging from May 1, 2014, through August 30, 2015, within the Indian Health Service-JVN program, which serves American Indian and Alaska Native communities at 97 sites across 25 states, were evaluated. All retinal images were graded using a standardized validated protocol in a centralized reading center. MAIN OUTCOME MEASURES: Ungradable rate for DR and diabetic macular edema (DME). RESULTS: The ungradable rate per patient for DR and DME was significantly lower with UWF imaging compared with NMFP (DR, 2.8% vs. 26.9% [P < 0.0001]; DME, 3.8% vs. 26.2% [P < 0.0001]). Identification of eyes with either DR or referable DR (moderate nonproliferative DR or DME or worse) was increased using UWF imaging from 11.7% to 24.2% (P < 0.0001) and from 6.2% to 13.6% (P < 0.0001), respectively. In eyes with DR imaged with UWF imaging (n = 3926 eyes of 2402 patients), the presence of predominantly peripheral lesions suggested a more severe level of DR in 7.2% of eyes (9.6% of patients). CONCLUSIONS: In a large, widely distributed DR ocular telehealth program, as compared with NMFP, nonmydriatic UWF imaging reduced the number of ungradable eyes by 81%, increased the identification of DR nearly 2-fold, and identified peripheral lesions suggesting more severe DR in almost 10% of patients, thus demonstrating significant benefits of this imaging method for large DR teleophthalmology programs.

RETINAL ANGIOMATOUS PROLIFERATION: A Quantitative Analysis of the Fundoscopic Features of the Fellow Eye.

PURPOSE: To quantitatively analyze and compare the fundoscopic features between fellow eyes of retinal angiomatous proliferation and typical exudative age-related macular degeneration and to identify possible predictors of neovascularization. METHODS: Retrospective case-control study. Seventy-nine fellow eyes of unilateral retinal angiomatous proliferation (n = 40) and typical exudative age-related macular degeneration (n = 39) were included. Fundoscopic features of the fellow eyes were assessed using digital color fundus photographs taken at the time of diagnosis of neovascularization in the first affected eye. Grading was performed by two independent graders using RetmarkerAMD, a computer-assisted grading software based on the International Classification and Grading System for age-related macular degeneration. RESULTS: Baseline total number and area (square micrometers) of drusen in the central 1,000, 3,000, and 6,000 µm were considerably inferior in the fellow eyes of retinal angiomatous proliferation, with statistically significant differences (P < 0.05) observed in virtually every location (1,000, 3,000, and 6,000 µm). A soft drusen (≥125 µm) area >510,196 µm2 in the central 6,000 µm was associated with an increased risk of neovascularization (hazard ratio, 4.35; 95% confidence interval [1.56-12.15]; P = 0.005). CONCLUSION: Baseline fundoscopic features of the fellow eye differ significantly between retinal angiomatous proliferation and typical exudative age-related macular degeneration. A large area (>510,196 µm2) of soft drusen in the central 6,000 µm confers a significantly higher risk of neovascularization and should be considered as a phenotypic risk factor.

Applying a verification classification system to medical and forensic digital images.

This paper gives the definitions, and describes the application, of a classification system for photographic veracity relating to medical and forensic images. It uses practical examples to demonstrate the reasons and parameters for using such a system. It also calls for continued discussion on what the limits and extent should be for the parameters. A debate encompassing the advantages and disadvantages of manipulating photographs goes back well before the introduction of digital imaging and has continued since the introduction of photography in the mid-nineteenth century. This author's first foray dates back twenty years (1). The introduction of digital imaging only gave the debate new impetus and new problems to solve (2), some of which might be resolved as suggested in this paper.

Disability in physical education textbooks: an analysis of image content.

The aim of this paper is to show how images of disability are portrayed in physical education textbooks for secondary schools in Spain. The sample was composed of 3,316 images published in 36 textbooks by 10 publishing houses. A content analysis was carried out using a coding scheme based on categories employed in other similar studies and adapted to the requirements of this study with additional categories. The variables were camera angle, gender, type of physical activity, field of practice, space, and level. Univariate and bivariate descriptive analyses were also carried out. The Pearson chi-square statistic was used to identify associations between the variables. Results showed a noticeable imbalance between people with disabilities and people without disabilities, and women with disabilities were less frequently represented than men with disabilities. People with disabilities were depicted as participating in a very limited variety of segregated, competitive, and elite sports activities.

Glaucoma Expert-Level Detection of Angle Closure in Goniophotographs With Convolutional Neural Networks: The Chinese American Eye Study.

PURPOSE: To compare the performance of a novel convolutional neural network (CNN) classifier and human graders in detecting angle closure in EyeCam (Clarity Medical Systems, Pleasanton, California, USA) goniophotographs. DESIGN: Retrospective cross-sectional study. METHODS: Subjects from the Chinese American Eye Study underwent EyeCam goniophotography in 4 angle quadrants. A CNN classifier based on the ResNet-50 architecture was trained to detect angle closure, defined as inability to visualize the pigmented trabecular meshwork, using reference labels by a single experienced glaucoma specialist. The performance of the CNN classifier was assessed using an independent test dataset and reference labels by the single glaucoma specialist or a panel of 3 glaucoma specialists. This performance was compared to that of 9 human graders with a range of clinical experience. Outcome measures included area under the receiver operating characteristic curve (AUC) metrics and Cohen kappa coefficients in the binary classification of open or closed angle. RESULTS: The CNN classifier was developed using 29,706 open and 2,929 closed angle images. The independent test dataset was composed of 600 open and 400 closed angle images. The CNN classifier achieved excellent performance based on single-grader (AUC = 0.969) and consensus (AUC = 0.952) labels. The agreement between the CNN classifier and consensus labels (κ = 0.746) surpassed that of all non-reference human graders (κ = 0.578-0.702). Human grader agreement with consensus labels improved with clinical experience (P = 0.03). CONCLUSION: A CNN classifier can effectively detect angle closure in goniophotographs with performance comparable to that of an experienced glaucoma specialist. This provides an automated method to support remote detection of patients at risk for primary angle closure glaucoma.

Evaluation of the age-related eye disease study clinical lens grading system AREDS report No. 31.

PURPOSE: To examine the grading (interrater) reliability of the Age-Related Eye Disease Study (AREDS) Clinical Lens Grading System (ARLNS). DESIGN: Evaluation of diagnostic test or technology. PARTICIPANTS: One hundred fifty volunteers (284 eyes). METHODS: Participants with lens opacities of varying severity were independently graded at the slit lamp for cataract severity by 2 examiners (retinal or anterior segment specialists) using the ARLNS, which employs 3 standard photographs of increasing severity for classifying each of the 3 major types of opacity. Lens photographs were taken and graded at a reading center using the more detailed AREDS System for Classifying Cataracts from photographs. MAIN OUTCOME MEASURES: The Pearson correlation, weighted-kappa, and limits-of-agreement statistics were used to assess the interrater agreement of the gradings. RESULTS: Examinations were performed on 284 lenses (150 participants). Tests of interrater reliability between pairs of clinicians showed substantial agreement between clinicians for cortical and posterior subcapsular opacities and moderate agreement for nuclear opacities. A similar pattern and strength of agreement was present when comparing scores of retinal versus anterior segment specialists. Interrater agreement between clinical and reading center gradings was not as great as inter-clinician agreement. CONCLUSIONS: Interrater agreements were in the moderate to substantial range for the clinical assessment of lens opacities. Inherent differences in cataract classification systems that rely on slit lamp vs photographic assessments of lens opacities may explain some of the disagreement noted between slit lamp and photographic gradings. Given the interrater reliability statistics for clinicians and the simplicity of the grading procedure, ARLNS is presented for use in studies requiring a simple, inexpensive method for detecting the presence and severity of the major types of lens opacities. FINANCIAL DISCLOSURE(S): The authors have no proprietary or commercial interest in any of the materials discussed in this article.

Caption-based topical descriptors for microscopic images as published in academic papers.

BACKGROUND: Visual findings summarized in the figures and tables of academic papers are invaluable sources for biomedical researchers. Captions associated with the visual findings are often neglected while retrieving biomedical images in published academic papers. OBJECTIVES: This study is to assess caption-based topical descriptors for microscopic images of breast neoplasms, as published in academic papers retrieved through the PubMed Central database. METHOD: Human indexers as well as an automatic keyword finder called TAPoR generated the topical descriptors from collected captions. The study then compared the human-generated descriptors to machine-generated descriptors. Finally, a set of core descriptors was developed from both sets and automatically mapped into the Unified Medical Language System's (UMLS) Metathesaurus through a MetaMap Transfer engine. RESULTS: Major topical descriptors included histologic disease names, laboratory procedures, genetic functions and components. Human indexers provided more relevant descriptors than TAPoR. The UMLS Metathesaurus identified several semantic types including Indicator, Reagent, or Diagnostic Aid; Organic Chemical; Laboratory Procedure; Spatial Concept; Qualitative Concept; and Quantitative Concept. DISCUSSION: The findings suggest that caption-based descriptors can complement title or abstract-based literature indexing for figure image retrieval in articles. With respect to forming a metadata framework for online microscopic image description, the semantic types can be used as a core metadata set. In this regard, this finding can be used for standardising a microscopic image description protocol to train medical students. CONCLUSIONS: It is incumbent upon libraries and other information agencies to promote and maintain an interest in the opportunities and challenges associated with biomedical imaging.

Hierarchical Approach to Classify Food Scenes in Egocentric Photo-Streams.

Recent studies have shown that the environment where people eat can affect their nutritional behavior [1]. In this paper, we provide automatic tools for personalized analysis of a person's health habits by the examination of daily recorded egocentric photo-streams. Specifically, we propose a new automatic approach for the classification of food-related environments, that is able to classify up to 15 such scenes. In this way, people can monitor the context around their food intake in order to get an objective insight into their daily eating routine. We propose a model that classifies food-related scenes organized in a semantic hierarchy. Additionally, we present and make available a new egocentric dataset composed of more than 33 000 images recorded by a wearable camera, over which our proposed model has been tested. Our approach obtains an accuracy and F-score of 56% and 65%, respectively, clearly outperforming the baseline methods.

Blind camera fingerprinting and image clustering.

Previous studies have shown how to "fingerprint" a digital camera given a set of images known to come from the camera. A clustering technique is proposed to construct such fingerprints from a mixed set of images, enabling identification of each image's source camera without any prior knowledge of source.

Accuracy of trained rural ophthalmologists versus non-medical image graders in the diagnosis of diabetic retinopathy in rural China.

BACKGROUND/AIMS: To determine the diagnostic accuracy of trained rural ophthalmologists and non-medical image graders in the assessment of diabetic retinopathy (DR) in rural China. METHODS: Consecutive patients with diabetes mellitus were examined from January 2014 to December 2015 at 10 county-level facilities in rural Southern China. Trained rural ophthalmologists performed a complete eye examination, recording diagnoses using the UK National Diabetic Eye Screening Programme (NDESP) classification system. Two field, mydriatic, 45° digital photographs were made by nurses using NDESP protocols and graded by trained graders with no medical background using the NDESP system. A fellowship-trained retina specialist graded all images in masked fashion and served as reference standard. RESULTS: Altogether, 375 participants (mean age 60±10 years, 48% men) were examined and 1277 images were graded. Grader sensitivity (0.82-0.94 (median 0.88)) and specificity (0.91-0.99 (median 0.98)), reached or exceeded NDESP standards (sensitivity 80%, specificity 95%) in all domains except specificity detecting any DR. Rural ophthalmologists' sensitivity was 0.65-0.95 (median 0.66) and specificity 0.59-0.95 (median 0.91). There was strong agreement between graders and the reference standard (kappa=0.84-0.87, p<0.001) and weak to moderate agreement between rural doctors and the reference (kappa=0.48-0.64, p<0.001). CONCLUSION: This is the first study of diagnostic accuracy in DR grading among non-medical graders or ophthalmologists in low-income and middle-income countries. Non-medical graders can achieve high levels of accuracy, whereas accuracy of trained rural ophthalmologists is not optimal.

Lash ptosis in congenital and acquired blepharoptosis.

OBJECTIVE: To determine the prevalence of lash ptosis (LP) in eyes with congenital and acquired blepharoptosis. METHODS: We retrospectively graded photographs of 228 eyes from 174 patients with congenital or acquired blepharoptosis for LP. We used a 4-point rating scale for LP, in which 0 indicates no LP; 1, minimal; 2, moderate; and 3, severe. A prospective evaluation of LP in 30 eyes from 15 patients without blepharoptosis (control eyes) was also performed. RESULTS: A total of 107 eyes (in 87 patients) demonstrated congenital blepharoptosis and 121 eyes (in 87 patients) had acquired blepharoptosis. A moderate to severe rating of LP (rating, >/= 2) occurred in 60.7% of eyes with congenital blepharoptosis, 28.9% of eyes with acquired blepharoptosis, and 6.7% of control eyes. Lash ptosis (rating, >/= 1) was present in 91.6% of eyes with congenital blepharoptosis, 83.5% of eyes with acquired blepharoptosis, and 33.3% of control eyes. The mean LP rating was 2.1 for eyes with congenital blepharoptosis, 1.3 for eyes with acquired blepharoptosis, and 0.6 for control eyes. CONCLUSIONS: Lash ptosis was common in the patients with blepharoptosis. Moderate to severe LP occurred more commonly in all forms of blepharoptosis compared with normal eyes, with more frequent and more severe LP demonstrated in eyes with congenital blepharoptosis.

Information loss in 2 bleb grading systems.

PURPOSE: To assess the effectiveness of transferring descriptive information from bleb photographs to 2 recently described bleb grading systems: the Moorfields Bleb Grading System (MBGS) and the Indiana Bleb Appearance Grading Scale (IBAGS). METHODS: Two experienced observers graded 51 clinical bleb photographs with a wide range of appearances using both the MBGS and IBAGS bleb grading systems in random order. Grading scores from the 2 observers were averaged, and these numbers used by a third investigator, who did not view the original photographs, to generate 102 sketched representations of the blebs. The sketches were labeled randomly, and 1 month later presented individually in random order, to mask which grading system was used as source data for each drawing. MAIN OUTCOME MEASURES: The original graders then used an arbitrary 1-5 scale to rate congruity between sketches and photographs for vascularity and morphology features, and overall agreement of the bleb sketches. RESULTS: For both the IBAGS and MBGS, interobserver agreement between the Congruity Scores (CS) of the 2 masked graders was excellent, ranging between 92% and 98% for each parameter. Overall CS results were 3.2 (good-very good) for IBAGS and 4.1 (very good-excellent) for MBGS. Vascularity CS scores from IBAGS were 3.0 (good) and those from morphology agreement averaged 3.5 (good-very good). For the MBGS, the respective results were 3.9 (good-very good) and 4.1 (very good-excellent), respectively. Photographic quality (P=0.012) and presence of a limbus-based conjunctival flap scar (P=0.012) had an influence on CS scores from IBAGS but not from MBGS. CONCLUSIONS: Both the IBAGS and MBGS produced acceptable agreement ratings between the sketches derived from grading system data and the original bleb photographs. These grading systems seem to adequately represent the blebs that are being encoded, without significant information loss from the simplification and translation process. The MBGS tended to have higher CS, and may be less influenced by photograph quality and bleb type, suggesting that bleb photographs may be best encoded for statistical analysis in clinical studies using this system.

The grey fovea sign of macular oedema or subfoveal fluid on non-stereoscopic fundus photographs.

PURPOSE: To describe the grey fovea sign of fovea-involving macular oedema or subretinal fluid accumulation in red-free fundus photography. METHODS: A test set of 91 digital fundus photographs of good quality from 100 consecutive eyes in 72 patients with diabetic retinopathy or central serous chorioretinopathy was composed by one of the investigators and evaluated by four masked observers. The photographs were graded as to whether a normal dark fovea was present or absent. The reference method was foveal thickness measurement using optical coherence tomography (OCT). RESULTS: Eyes graded as having a grey fovea on fundus photographs (n = 67) had a median foveal thickness of 279 µm (interquartile range 130 µm), whereas eyes graded as having a normal dark fovea (n = 24) had a median foveal thickness of 238 µm (interquartile range 44.5 µm, p = 0.025). CONCLUSION: The absence of a dark fovea on red-free greyscale fundus photographs is a sign of foveal thickening or detachment that can be reliably recognized by graders when using OCT as the reference diagnostic method. Awareness of the grey fovea sign may facilitate fundus photographic screening for maculopathy because its absence is a reliable sign that no foveal oedema or detachment is present. Its presence may indicate that such conditions are present, but it can also be induced by various other conditions, for which reason it should prompt further investigations by OCT.

Visual acuity thresholds for cataract surgery and the changing Australian population.

OBJECTIVE: To examine the relation between visual acuity thresholds for cataract surgery for the changing Australian population. METHODS: Population-based 5-year incidence data for 2594 people were used to calculate age-specific prevalence of incident cataract and existing cataract backlog. The cataract surgery rates (CSRs) per million people were estimated for different acuity thresholds for the Australian population at different times. RESULTS: In 2005, after adjusting by age, the estimated CSRs for Australia were 1470, 3110, 5080, and 6440 for visual acuities of less than 6/60, less than 6/18, less than 6/12, and less than 6/9, respectively, whereas the actual CSR was 9000. Although the acuity threshold had the greatest impact, the CSR for less than 6/12 increased 3-fold from 2300 in 1950 to 7210 in 2020 because of demographic changes. CONCLUSIONS: Demographic changes, with an increasing proportion of older people, will continue to increase the need for cataract surgery, as will a reduction in the visual impairment threshold. Plans for the future for eye care services should take these factors into account.

Agreement between novice and experienced trachoma graders improves after a single day of didactic training.

BACKGROUND/AIMS: Prevalence estimates and treatment decisions for trachoma are based entirely on ocular clinical examination. The aim of the current study is to demonstrate that ophthalmic assistants can be trained and certified to provide trachoma grading within a single day. METHODS: Conjunctival photographs from an area with endemic trachoma were randomised into two sets of 60 cases. Photographs were graded for trachomatous inflammation-follicular (TF) and trachomatous inflammation-intense (TI) by three experienced graders. Inter-rater reliability of eight ophthalmic assistants and three experienced graders were compared before and after training. RESULTS: The mean κ agreement between the ophthalmic assistants and the consensus grades of the experienced graders for TF was 0.38 (95% CI 0.18 to 0.58) before training, and increased to 0.60 (95% CI 0.42 to 0.78) after training (p=0.07). The mean κ agreement for TI was 0.16 (95% CI 0.02 to 0.30) before training, and increased to 0.39 (95% CI 0.20 to 0.58) after training (p=0.02). CONCLUSION: A single day of training improves agreement between prospective and experienced trachoma graders, and provides the basis for certification of workers who are able to accurately grade trachoma and generate reliable prevalence estimates.