Leveraging ResNet50 With Swin Attention for Accurate Detection of OCT Biomarkers Using Fundus Images
Diabetes can impact the retina and cause a decline in vision for patients as a result of Diabetic Retinopathy (DR). Diabetic Macular Edema (DME) is a complication that results from the chronic damage to the tiny blood vessels of the retina that arises in the non-proliferative stage of DR (NPDR) but...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10897982/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Diabetes can impact the retina and cause a decline in vision for patients as a result of Diabetic Retinopathy (DR). Diabetic Macular Edema (DME) is a complication that results from the chronic damage to the tiny blood vessels of the retina that arises in the non-proliferative stage of DR (NPDR) but can also be present in proliferative DR (PDR) and potentially leading to vision loss. The duration of diabetes in patients affects both the prevalence and incidence of macular edema, as well as the progression of retinopathy. Therefore, regular screening of diabetes patients for the early detection of retinal abnormalities is essential to prevent the development and progression of DR and DME. The proposed model predicts DME-associated biomarkers, typically identified in Optical Coherence Tomography (OCT), using 2D fundus images. These biomarkers include center-involved diabetic macular edema (ci-DME), neurosensory detachment (NSD), Intraretinal fluid (IRF), disorganization of the retinal inner layers (DRIL), hyperreflective foci (HRF), and disruptions in the inner segment/outer segment (IS/OS) junction, utilizing 2D fundus images. The model integrates the feature extraction capability of ResNet50 with the spatial structural domain knowledge provided by the Swin attention augmentation layer. 2D fundus image datasets were collected to train and evaluate the model. In two distinct datasets, the model achieved a validation accuracy of 85.7% (95% CI: 81.6–90.6%) and 89.5% (95% CI: 85.6–93.4%), Cohen’s Kappa of 0.68 (95% CI: 0.61–0.77) and 0.75 (95% CI: 0.67–0.82), sensitivity of 88.6% (95% CI: 85.6–92.1%) and 79.6% (95% CI: 70.6–85.4%), specificity of 79.6% (95% CI: 70.3–88.9%) and 93.7% (95% CI: 90.7–96.8%), respectively, with an overall validation accuracy of 87%. The proposed model helps in identifying the DME-associated biomarkers, using 2D fundus images making it a promising tool for detecting and assessing DME-related features. |
|---|---|
| ISSN: | 2169-3536 |