Abstract
Diabetic retinopathy (DR) diagnosis methods in the literature are usually criticized as being limit in diagnosing DR-related features or being lack of interpretability. To deal with these issues, this paper investigates the feasibility of diagnosing both DR severity levels and the presence of DR-related features in a two-step procedure. Specifically, this paper first analyzes the quality of annotations in DR grading by measuring inter-grader variability. Cosine similarity is considered to evaluate the inter-grader variability of the presence of DR-related features, and quadratic weighted Cohen's kappa is employed to assess the inter-grader variability of DR severity levels. Next, different annotation methods as follows are compared to DR severity prediction performance using logistic regression: 1) single annotations by single grader (SASG); 2) single annotations from multiple graders (SAMG); 3) multiple annotations by voting (MAV); and 4) double annotations with adjudication of disagreement (DAAD). Based on the comparison results, the feasibility of diagnosing both DR severity and features is investigated. In the experiments, 1589 fundus images graded by three retinal specialists and four general ophthalmologists are considered. The results demonstrate that retinal specialists are more consistent than general ophthalmologists in grading both the presence of DR-related features and DR severity. The SASG and MAV should be avoided if possible while the DAAD is the good option when prediction performance is the highest priority and the SAMG is especially beneficial when both prediction performance and grading costs are considered. The upper limit performance of DR severity prediction gets accuracy 95.6% and kappa 0.962. When DR-related feature prediction achieves average cosine similarity 0.823, it is potential to get accuracy 91.2% and kappa 0.905 for DR severity prediction in real applications. These results together suggest the potential of diagnosis of both DR severity and the presence of DR-related features in a two-step procedure.
Highlights
Diabetic retinopathy (DR) is the leading cause of preventable blindness among working-aged adults in the world [1]
The mean of average cosine similarity is 0.349 for general ophthalmologist pairs, which is much lower than 0.681 for retinal specialist pairs. These results indicate that the gradings from retinal specialists tend to be more accurate than those from general ophthalmologists
This study investigated the feasibility of diagnosing both DR severity levels and the presence of DR features by a two-step procedure
Summary
Diabetic retinopathy (DR) is the leading cause of preventable blindness among working-aged adults in the world [1]. It is estimated that 35% of all patients with diabetes mellitus suffer from DR [2], [3]. The risk of DR increases as the longer a person has diabetes mellitus. According to World Health Organization, DR is estimated to affect more than 77% of the patients who have had diabetes 20 years and more [4]. Fundus photography is one of the most commonly used imaging technique for the diagnosis of DR in the retina [5]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
