Retinal grading stops reading one raw photo: separate lesion and vessel paths reach 93.32% on five-grade diabetic retinopathy staging
Synopsis
Retinal disease grading has mostly relied on raw color fundus photographs, where low-contrast lesions and thin vessels are easily lost to illumination and noise. F2D-RetinaNet sends an enhanced green channel and a vessel map through two separate branches and fuses them into one grading decision. On APTOS 2019 five-grade diabetic retinopathy staging it reaches 93.32% accuracy, 3.25 points above the RGB-only baseline, and 87.14% on the external DREAM-RFI set after overlapping images were excluded.
Interpretation
The grading decision now reads two images at once: an enhanced green channel and a vessel map extracted from the same fundus photograph. Most earlier models took only the raw RGB fundus image, so early lesions in low-contrast regions and fine vessels were hard to present together. On APTOS 2019 five-grade staging the dual-branch configuration reached 93.32% accuracy against 90.07% for the RGB-only baseline, a 3.25-point gap.
Each image follows its own branch: the enhanced image through DenseNet-169 and the vessel map through ResNet-50, with features concatenated and passed to a fusion module. The enhancement stage replaces conventional smoothing with fractional anisotropic diffusion on the green channel, followed by CLAHE, bilateral filtering and unsharp masking. On the 200-image FIVES test split the full enhancement pipeline significantly exceeded a CLAHE-only baseline in contrast-to-noise ratio, and swapping conventional for fractional anisotropic diffusion added about 1.5% accuracy.
The vessel map comes from FHAU-Net, which stacks eight fractional-order Hessian vesselness responses with the enhanced image into a 9-channel input for an attention-gated U-Net. A single fractional order covers only part of the vessel scale range, so thin and large vessels are hard to capture together. On a unified test set pooling 913 images from DRIVE, STARE, CHASE_DB1, HRF and FIVES, Dice was 0.8987, sensitivity 0.9084 and specificity 0.9876; leave-one-dataset-out Dice ranged from 0.7411 to 0.8009.
Perspective
Evaluation covers APTOS 2019 detection and five-grade staging, FIVES four-class disease classification, and the 2,156-image DREAM-RFI external set after overlapping images were excluded. Clinical feasibility was observed only on 19 retrospective fundus images from SMS Hospital Jaipur, 16 of which were classified correctly. The authors call for larger, prospectively collected clinical data for validation.
The authors note that annotation protocols were not harmonized across datasets, and inter-observer disagreement caps achievable cross-dataset segmentation performance. Mild and severe NPDR are more often misclassified because of class imbalance, and mild NPDR is easily mistaken for no retinopathy. Grad-CAM maps are shown for basic transparency only, without quantitative faithfulness analysis or expert ophthalmologist assessment.