Abstract
Background
Distinguishing between active and inactive level 4 diabetic retinopathy (DR) cases on the International Classification of Diabetic Retinopathy (ICDR) is diagnostically imperative in order to treat patients with a sight-threatening retinal disease.
Methods
We included 2,816 retinal images from patients with type 1 or 2 diabetes that were citizens of the Funen Region and attended the Danish screening program for DR. The retinal images, classified as ICDR level 4, originated from The Funen Diabetes Database. A certified grader manually dichotomized images into either inactive (n = 2,196) or active (n = 620) PDR as a gold standard evaluation. Afterwards, we tested our pre-established DL model, which is capable of segmenting eight different DR lesions and classifying DR according to the ICDR scale (0 to 4), on the same set of retinal images. We considered a retinal image graded as active PDR if the DL model marked at least one lesion as a neovascularization or a pre-retinal bleeding, and then calculated the sensitivity and specificity using the certified grader as the reference.
Results
The certified grader identified active PDR in 620 cases (22%) and the DL model identified 1,756 cases (62%) out of a total of 2,816 retinal images with ICDR-level 4. Testing the DL model for detection of active PDR resulted in a sensitivity of 86%, a specificity of 44%, a positive predictive value of 31% and a negative predictive value of 92%.
Conclusion
The DL model demonstrated a noteworthy sensitivity of 86%, emphasizing its potential to identify a large proportion of the true positive cases. However, the lower specificity of 44% indicates a higher rate of false positive cases, but the DL model did on the other hand demonstrate a negative predictive value of 92%, indicating its proficiency in correctly identifying cases without active PDR.
There is a need for resegmentation of neovascularizations and preretinal bleedings followed by retraining of the DL model to improve the overall performance in classifying inactive versus active PDR.
Distinguishing between active and inactive level 4 diabetic retinopathy (DR) cases on the International Classification of Diabetic Retinopathy (ICDR) is diagnostically imperative in order to treat patients with a sight-threatening retinal disease.
Methods
We included 2,816 retinal images from patients with type 1 or 2 diabetes that were citizens of the Funen Region and attended the Danish screening program for DR. The retinal images, classified as ICDR level 4, originated from The Funen Diabetes Database. A certified grader manually dichotomized images into either inactive (n = 2,196) or active (n = 620) PDR as a gold standard evaluation. Afterwards, we tested our pre-established DL model, which is capable of segmenting eight different DR lesions and classifying DR according to the ICDR scale (0 to 4), on the same set of retinal images. We considered a retinal image graded as active PDR if the DL model marked at least one lesion as a neovascularization or a pre-retinal bleeding, and then calculated the sensitivity and specificity using the certified grader as the reference.
Results
The certified grader identified active PDR in 620 cases (22%) and the DL model identified 1,756 cases (62%) out of a total of 2,816 retinal images with ICDR-level 4. Testing the DL model for detection of active PDR resulted in a sensitivity of 86%, a specificity of 44%, a positive predictive value of 31% and a negative predictive value of 92%.
Conclusion
The DL model demonstrated a noteworthy sensitivity of 86%, emphasizing its potential to identify a large proportion of the true positive cases. However, the lower specificity of 44% indicates a higher rate of false positive cases, but the DL model did on the other hand demonstrate a negative predictive value of 92%, indicating its proficiency in correctly identifying cases without active PDR.
There is a need for resegmentation of neovascularizations and preretinal bleedings followed by retraining of the DL model to improve the overall performance in classifying inactive versus active PDR.
Originalsprog | Engelsk |
---|---|
Tidsskrift | European Journal of Ophthalmology |
Vol/bind | 34 |
Udgave nummer | 1 Suppl. |
ISSN | 1120-6721 |
Status | Udgivet - 2024 |
Begivenhed | EAsDEC - Varighed: 30. maj 2024 → 1. jun. 2024 |
Konference
Konference | EAsDEC |
---|---|
Periode | 30/05/2024 → 01/06/2024 |