Adaptive Retinal Image Enhancement Methods

• Adaptive retinal image enhancement is a class of techniques that adjust local contrast and illumination to make diagnostically relevant structures, like vessels and lesions, more visible.
• These methods combine classical approaches such as illumination-aware contrast stretching with modern learning-based models to balance photometric improvement and clinical fidelity.
• Applications include preprocessing for vessel segmentation, diabetic retinopathy grading, and enhancing both color fundus and ultra-wide-field imaging for improved diagnostic analysis.

Searching arXiv for recent and foundational papers on retinal image enhancement and related fundus/UWF enhancement methods. Adaptive retinal image enhancement denotes the class of methods that modify retinal photographs so that diagnostically relevant structures become more visible or more separable for downstream analysis. Across the cited literature, the target modalities include conventional color fundus photography, retinal vessel images, and ultra-wide-field retinal imaging, while the targeted degradations include nonuniform illumination, low and spatially varying contrast, blur, noise, artifacts, and color inconsistency. The field spans handcrafted illumination-aware contrast transforms, fuzzy and histogram-based hybrids, Retinex- and dehazing-inspired formulations, unpaired generative translation, self-supervised frequency-aware restoration, and segmentation-driven feature enhancement. A recurring theme is that “adaptive” enhancement is not a single mechanism but a family of strategies in which the transformation depends on image brightness, local contrast, frequency content, retinal structure, or the geometry of clinically related samples in feature space (Bandara et al., 2018, Gaudio et al., 2020, Shaout et al., 25 Feb 2025, Liao et al., 27 Aug 2025, Xiong et al., 19 Jan 2026, Dong et al., 2024).

1. Historical framing and problem definition

Retinal enhancement is typically motivated as a preprocessing stage for vessel segmentation, lesion analysis, diabetic retinopathy grading, or visual interpretation. In one vessel-segmentation line, contrast enhancement is described as “one of the crucial steps” because segmentation reliability depends on consistency of contrast over the image, especially in the presence of thin vessels, optic-disc interference, and the fovea (Bandara et al., 2018). Another unsupervised segmentation pipeline similarly emphasizes non-uniform illumination, poor global and local contrast, noise amplification, weak vessel-edge visibility, and optic-disc interference as major causes of segmentation failure (Chalakkal et al., 2017). A later fundus-enhancement study states that retinal fundus images often require enhancement before analysis because segmentation models are sensitive to image appearance, and it reports Dice score increases over a no-enhancement baseline by as much as 0.491 in a challenging multi-task setting (Gaudio et al., 2020).

The scope of the problem has widened with modality. Ultra-wide-field retinal imaging introduces severe peripheral illumination inhomogeneity, blur, and red/green laser-induced color behavior that conventional fundus methods do not explicitly address (Liao et al., 27 Aug 2025). Unpaired enhancement for color fundus photography further reframes the task as improving degraded images while preserving disease-relevant structure, because global distribution matching can disperse clinically related samples and blur disease-class boundaries (Xiong et al., 19 Jan 2026). This suggests that adaptive retinal enhancement is best understood as a task-conditioned image transformation problem in which photometric improvement must be balanced against structural and clinical fidelity.

2. Classical adaptive enhancement: local illumination, contrast, and vessel visibility

A representative classical formulation is SUACE, a “spatially adaptive linear contrast stretching method” integrated into an improved Tyler Coye segmentation pipeline (Bandara et al., 2018). Its local lower and upper bounds are defined by

$$a(x,y)=g(x,y)-\frac{d}{2}, \qquad b(x,y)=g(x,y)+\frac{d}{2},$$

where the low-frequency response is

$g(x,y)=I(x,y) * f(x,y),$

with Gaussian kernel

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$

The enhanced image is then obtained by piecewise linear stretching and clipping:

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$

In the reported experiments, the parameters were fixed empirically to $\sigma=7$ and $d=16$, and the downstream pipeline used $a_1=10$, $h=5$, $v=3$, and $a_2=50$ (Bandara et al., 2018). On STARE and DRIVE, this configuration achieved the best reported accuracy among the compared enhancement methods, with ACC values of 0.9489 and 0.9411, respectively (Bandara et al., 2018).

A distinct classical pipeline is based on luminance enhancement in CIELab, anisotropic diffusion filtering, weighted grayscale conversion, curvelet-domain edge enhancement, and optic-disc suppression (Chalakkal et al., 2017). The grayscale conversion is explicitly

$g(x,y)=I(x,y) * f(x,y),$0

and optic-disc removal uses

$g(x,y)=I(x,y) * f(x,y),$1

with a median low-pass filter of size $g(x,y)=I(x,y) * f(x,y),$2, followed by

$g(x,y)=I(x,y) * f(x,y),$3

$g(x,y)=I(x,y) * f(x,y),$4

and

$g(x,y)=I(x,y) * f(x,y),$5

The enhancement stages are followed by fuzzy C-means vessel extraction, and the method reports 95.18% segmentation accuracy on DRIVE (Chalakkal et al., 2017). The paper attributes robustness on difficult images to the preprocessing sequence rather than to the clustering stage alone.

These methods exemplify a first major meaning of adaptivity: enhancement parameters are made functions of local illumination, local background, or directional vessel-like detail rather than fixed global constants. A plausible implication is that these pipelines remain relevant when interpretability and lightweight computation are prioritized over learned feature hierarchies.

3. Hybrid contrast enhancement with global and local adaptation

A more explicit hybrid formulation appears in the fuzzy-based retinal contrast enhancement method developed for retinal vascular structure segmentation (Shaout et al., 25 Feb 2025). The image is processed in HLS color space, and the luminosity values are fuzzified into five linguistic classes—Very Dark, Dark, Medium, Bright, and Very Bright—using Gaussian membership functions

$g(x,y)=I(x,y) * f(x,y),$6

The memberships are defined as

$g(x,y)=I(x,y) * f(x,y),$7

$g(x,y)=I(x,y) * f(x,y),$8

$g(x,y)=I(x,y) * f(x,y),$9

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$0

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$1

Here the centers and spreads are controlled by the image-level brightness statistic $$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$2, so the fuzzy partition shifts with overall image luminosity (Shaout et al., 25 Feb 2025). The rule base is deliberately simple: IF input is Dark THEN output is Very Dark; IF input is Medium THEN output is Medium; IF input is Bright THEN output is Very Bright (Shaout et al., 25 Feb 2025). After inference, min–max normalization is applied in a typeset-corrupted but clearly intended form,

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$3

The fuzzy stage is then blended with CLAHE through

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$4

using

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$5

The method is evaluated on DRIVE against grayscale conversion, histogram equalization, FCE, and CLAHE, and the only reported quantitative result is a human survey involving ten individuals in which “both the FCE and FCE + CLAHE methods outperformed others, achieving a remarkable 88% as the preferred enhancement methods” (Shaout et al., 25 Feb 2025). The paper explicitly notes that this 88% is a preference statistic rather than segmentation accuracy or an image-quality metric (Shaout et al., 25 Feb 2025).

This hybrid scheme makes the distinction between global and local adaptivity explicit. The FCE component is globally adaptive because its membership functions depend on $$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$6, while CLAHE contributes local histogram equalization over subregions (Shaout et al., 25 Feb 2025). This suggests that one persistent design pattern in retinal enhancement is the coupling of image-level photometric adaptation with region-level contrast equalization.

4. Retinex, dehazing reinterpretation, and illumination–reflectance modeling

Retinex-style decomposition and related image-formation models provide another major axis of adaptive retinal enhancement. One fundus-specific study proposes a “pixel color amplification theory” and a family of retinal enhancement methods derived from “a novel re-interpretation of the image distortion model underlying dehazing theory,” relating three existing dehazing priors and a novel fourth prior (Gaudio et al., 2020). The abstract does not provide the equations, but it explicitly states that the resulting family includes whole-image brightening, whole-image darkening, and a novel derivation of Unsharp Masking, and that enhancement composition can act as class balancing on unbalanced and difficult data (Gaudio et al., 2020).

A UWF-specific self-supervised method makes the Retinex structure explicit through its Retinex-guided illumination compensation module, where

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$7

$$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$8

and under the assumption $$f(x,y)=\frac{1}{2\pi \sigma^2}\exp\!\left(-\frac{x^2+y^2}{2\sigma^2}\right).$$9,

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$0

with

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$1

as the illumination compensation ratio to be estimated by the network (Liao et al., 27 Aug 2025). The same paper adds a color preservation unit based on the 2D discrete wavelet transform with filters

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$2

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$3

and reconstructs fused features by

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$4

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$5

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$6

This formulation makes illumination correction adaptive to image-specific ratio estimation and frequency structure rather than to a fixed global remapping (Liao et al., 27 Aug 2025).

Related low-light literature that is not retina-targeted but is Retinex-relevant introduces additional adaptive ideas. A self-regularized Retinex variant preserves hue and saturation and enhances only the HSV value channel, using

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$7

$\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$8

with inverse illumination $\hat{I}(x,y)= \begin{cases} 0, & I(x,y)<a(x,y)\[4pt] k, & I(x,y)\ge b(x,y)\[4pt] \displaystyle \frac{I(x,y)-a(x,y)}{d}\,k, & \text{otherwise}. \end{cases}$9, and trains with

$\sigma=7$0

under

$\sigma=7$1

$\sigma=7$2

$\sigma=7$3

$\sigma=7$4

The paper includes no retinal experiments, but it suggests a route for color-preserving brightness-only enhancement that could be adapted for fundus imaging (Jiang et al., 2021). Similarly, a visual-perception low-light model decomposes bright and dark energy and area statistics into adaptive control variables $\sigma=7$5 and $\sigma=7$6; this is also not retinally validated, but it suggests that enhancement can be regulated by the spatial distribution of bright and dark content rather than by intensity alone (Li et al., 2020). These are best read as methodological context rather than direct retinal evidence.

5. Learning-based adaptive enhancement: unpaired translation, geometry preservation, and UWF restoration

Recent retinal enhancement methods move from handcrafted adaptive mappings to learned content-adaptive translation. CUNSB-RFIE formulates unpaired fundus enhancement as a Schrödinger Bridge problem between low-quality and high-quality retinal distributions (Dong et al., 2024). The core bridge formulation is

$\sigma=7$7

with the static entropy-regularized OT form

$\sigma=7$8

and conditional bridge state

$\sigma=7$9

Its final objective is

$d=16$0

with $d=16$1, $d=16$2, $d=16$3, $d=16$4, and $d=16$5 (Dong et al., 2024). The generator is U-Net-like and incorporates Dynamic Snake Convolution to preserve tortuous vessels; ablation on EyeQ shows PSNR/SSIM rising from 22.232/0.793 without DSC or regularization to 27.611/0.858 with DSC, PatchNCE, and SSIM (Dong et al., 2024). On downstream DRIVE vessel segmentation, the method reports ROC 0.918, PR 0.771, and SP 0.979, but it is weaker than OTEGAN on IDRiD lesion segmentation, which the paper attributes to smoothing of high-frequency lesion details (Dong et al., 2024). This directly illustrates an important controversy: enhancement that benefits vessel preservation may still underperform on punctate lesions.

SGW-GAN makes the preservation issue more explicit by arguing that enhancement should preserve intra-class geometry in a clinically meaningful feature space rather than only match the high-quality image distribution (Xiong et al., 19 Jan 2026). Its generator objective is

$d=16$6

with

$d=16$7

$d=16$8

$d=16$9

The SGW term is computed on frozen RETFound embeddings, with

$a_1=10$0

where $a_1=10$1 projections are used in training (Xiong et al., 19 Jan 2026). On EyeQ, SGW-GAN achieves SSIM 0.9035, PSNR 23.914, FID 18.141, ACC 0.7814, F1 0.7749, AUC 0.9297, and GW discrepancy 599.92, outperforming the listed baselines on FID, ACC, F1, and GW discrepancy while not winning on SSIM or PSNR (Xiong et al., 19 Jan 2026). Its ablation is particularly revealing: removing SGW improves SSIM and PSNR to 0.9288 and 24.417 but worsens GW discrepancy to 847.68 (Xiong et al., 19 Jan 2026). The paper interprets this as evidence that standard visual metrics can improve while clinical geometry deteriorates.

For UWF images, frequency-aware self-supervised learning combines frequency-decoupled deblurring and Retinex-guided illumination compensation (Liao et al., 27 Aug 2025). The deblurring module optimizes

$a_1=10$2

while illumination compensation uses

$a_1=10$3

with $a_1=10$4, $a_1=10$5, and $a_1=10$6 (Liao et al., 27 Aug 2025). The model is trained on 400 UWF images and evaluated on 434 images acquired with Optos 200Tx at resolution $a_1=10$7 and $a_1=10$8 field of view (Liao et al., 27 Aug 2025). Compared with the original images, it improves NIQE from 5.87 to 4.09, BRISQUE from 35.25 to 19.15, PIQE from 16.32 to 8.26, ACC from 57.13 to 69.49, and F1 from 48.24 to 64.07, corresponding to 12.3% ACC and 15.8% F1 gains in diabetic retinopathy grading (Liao et al., 27 Aug 2025). This is presented as the first dedicated UWF enhancement method (Liao et al., 27 Aug 2025).

6. Segmentation-driven enhancement and downstream task coupling

Not all adaptive retinal enhancement operates at the pixel level. HREFNet is explicitly a vessel-segmentation framework rather than a standalone image enhancer, but its internal design functions as adaptive feature enhancement specialized for retinal vessels (Ouyang et al., 18 Apr 2025). Its Multi-scale Retina Edge Fusion module computes

$a_1=10$9

$h=5$0

$h=5$1

$h=5$2

followed by

$h=5$3

and SE attention

$h=5$4

$h=5$5

$h=5$6

Its directional selective-scan mechanism uses

$h=5$7

$h=5$8

These operations enhance vessel-aligned edge and continuity information inside the network rather than modifying the input image (Ouyang et al., 18 Apr 2025).

On DRIVE, STARE, and CHASE_DB1, the method reports Dice/clDice/ACC/AUC/HD95 of 82.14 ± 1.53 / 82.40 ± 4.07 / 96.90 ± 0.24 / 98.56 ± 0.39 / 5.45, 76.29 ± 3.94 / 80.30 ± 3.65 / 96.70 ± 0.78 / 98.05 ± 0.69 / 17.10, and 80.46 ± 1.57 / 82.93 ± 2.50 / 97.36 ± 0.37 / 98.78 ± 0.35 / 10.34 ± 3.17, respectively (Ouyang et al., 18 Apr 2025). The paper interprets the gains as improved continuity and clearer boundaries in low-contrast regions (Ouyang et al., 18 Apr 2025). A plausible implication is that, for some retinal tasks, the most effective “enhancement” is not generic contrast modification but structure-aware representation enhancement coupled directly to the segmentation objective.

The same downstream perspective appears in earlier preprocessing papers. The dehazing-inspired color amplification study evaluates enhancement only through a challenging multi-task segmentation problem and reports the largest Dice gain—up to 0.491—on difficult and unbalanced data (Gaudio et al., 2020). The UWF method evaluates enhancement with both no-reference image-quality metrics and DR grading (Liao et al., 27 Aug 2025). SGW-GAN measures enhancement by DR grading and feature-space GW discrepancy rather than by restoration scores alone (Xiong et al., 19 Jan 2026). These examples indicate that retinal enhancement is increasingly judged by its effect on clinical or analytic tasks rather than by appearance alone.

7. Evaluation criteria, recurring trade-offs, and open directions

The literature uses heterogeneous evaluation protocols, and the mismatch among them is itself a substantive feature of the field. Classical vessel-oriented work reports TPR, FPR, and ACC through downstream segmentation on STARE and DRIVE (Bandara et al., 2018), or ACC, sensitivity, and specificity on DRIVE (Chalakkal et al., 2017). Fuzzy–CLAHE work relies primarily on a human preference survey rather than task metrics (Shaout et al., 25 Feb 2025). UWF enhancement uses NIQE, BRISQUE, PIQE, ACC, and F1 (Liao et al., 27 Aug 2025). CUNSB-RFIE uses PSNR and SSIM for enhancement and ROC, PR, sensitivity, specificity, and Jaccard for downstream segmentation (Dong et al., 2024). SGW-GAN reports SSIM, PSNR, FID, ACC, F1, AUC, and GW discrepancy (Xiong et al., 19 Jan 2026). This variety reflects different conceptions of enhancement quality: perceptual realism, structure preservation, topology preservation, and clinical fidelity are not equivalent objectives.

Several trade-offs recur. First, methods that sharpen or equalize local contrast can improve vessel visibility but amplify noise or distort bright structures. The fuzzy-based method explicitly notes that FCE alone can deteriorate fine vascular detail in high-luminosity regions, especially around the macular area, motivating the CLAHE blend (Shaout et al., 25 Feb 2025). Classical comparisons report box-shaped artifacts for CLAHE, random noise for local normalization, and black patches for wavelet- and contourlet-based methods (Bandara et al., 2018). Second, generative enhancement can improve realism while harming disease-related structure. SGW-GAN argues that global distribution alignment can blur class boundaries and shows that better SSIM and PSNR do not imply better geometry preservation (Xiong et al., 19 Jan 2026). Third, iterative probabilistic translation can preserve low-frequency structure yet oversmooth small lesions. CUNSB-RFIE achieves strong vessel results but underperforms OTEGAN on EX and HE lesion segmentation (Dong et al., 2024). Fourth, enhancement tuned for one structure may not transfer to another; the SUACE paper itself notes that future work could retune the method for fovea, optic disc, and retinal lesions (Bandara et al., 2018).

Three directions emerge from the cited work. One is modality-specific adaptation: UWF enhancement required a dedicated design with frequency-decoupled deblurring, Retinex-guided illumination compensation, and wavelet-based color preservation (Liao et al., 27 Aug 2025). A second is clinically structured regularization: SGW-GAN introduces sample-geometry preservation in a retinal feature space, while CUNSB-RFIE adds PatchNCE and SSIM to bridge learning (Xiong et al., 19 Jan 2026, Dong et al., 2024). A third is the migration from pixel-space enhancement to task-aware structural enhancement, exemplified by HREFNet’s edge and continuity modules (Ouyang et al., 18 Apr 2025). This suggests that adaptive retinal image enhancement is no longer reducible to contrast manipulation; it increasingly operates as a constrained transformation whose success is determined by whether retinal anatomy, pathology, and label geometry remain intact after enhancement.

In that sense, the field can be organized into three broad paradigms. The first comprises handcrafted illumination- and contrast-aware operators that adapt to local background or global brightness statistics (Bandara et al., 2018, Chalakkal et al., 2017, Shaout et al., 25 Feb 2025). The second comprises image-formation and decomposition models, including dehazing reinterpretations and Retinex-guided illumination compensation (Gaudio et al., 2020, Liao et al., 27 Aug 2025). The third comprises learned distributional or representation-level methods that adapt to retinal context, feature geometry, and downstream objectives (Dong et al., 2024, Xiong et al., 19 Jan 2026, Ouyang et al., 18 Apr 2025). The coexistence of these paradigms, and the lack of a single dominant evaluation regime, remains one of the defining characteristics of adaptive retinal image enhancement.