DR.INFO: Diabetic Retinopathy Diagnostic System

• DR.INFO is a diagnostic system for diabetic retinopathy that processes retinal images using morphological algorithms to detect key lesions.
• It employs rigorous image preprocessing and anatomical localization to enhance features like the optic disc and retinal lesions for accurate grading.
• The system integrates rule-based decision logic with clinical workflow solutions, achieving high performance metrics in automated DR screening.

Searching arXiv for the target paper and closely related diabetic retinopathy detection systems to ground the article. Algorithm-based diagnostic applications for diabetic retinopathy (DR) detection are computer-aided ophthalmic systems that analyze retinal images, identify anatomical landmarks and lesions, and return a screening, grading, or referral decision through a clinical interface. In the system described in "Algorithm-based diagnostic application for diabetic retinopathy detection," DR is treated as a growing health problem worldwide and a leading cause of visual impairment and blindness, especially among working people aged 20–65; the proposed pipeline processes hand-held ophthalmoscopic images, uses morphological algorithms to identify the optic disc and lesions characteristic of DR, and culminates in a graphical client-server application that uploads retinal images from cooperating ophthalmology offices for automated analysis and diagnosis (Cisek et al., 2023). The broader literature places this approach within a spectrum ranging from morphology- and threshold-based lesion extraction to Mask R-CNN instance segmentation, multimodal OCT/OCTA/LSO fusion, interactive human-in-the-loop grading, and referral-oriented responsible AI systems (Shenavarmasouleh et al., 2021).

1. Clinical scope and imaging targets

Diabetic retinopathy is a microvascular complication of diabetes mellitus in which lesions such as microaneurysms, hemorrhages, and exudates appear in the retina. In the target application, these lesion classes are explicitly operationalized as the image signatures on which the diagnosis is based, and automated detection is presented as a means to improve the efficiency of early disease detection and reduce diabetes-related visual impairment (Cisek et al., 2023).

The imaging substrate in the core system is the hand-held ophthalmoscope image. The evaluation set consists of 647 hand-held ophthalmoscope images from Optomed Smartscope® PRO, pre-labeled by ophthalmologists, indicating a deployment scenario centered on portable retinal imaging rather than only fixed-table fundus cameras (Cisek et al., 2023). Related work shows that the same diagnostic problem is also approached with 3-D structural OCT, 3-D OCTA, and 2-D LSO acquired simultaneously on a PLEX® Elite 9000, where the objective is automatic detection of proliferative DR rather than binary DR/no-DR screening (Li et al., 2023).

This modality variation is clinically significant because it changes both the observable features and the classification target. Fundus-image pipelines emphasize visible lesions and anatomical exclusion zones; OCT/OCTA/LSO fusion emphasizes retinal layer morphology, microvascular perfusion, and en face reflectance. This suggests that “algorithm-based” DR detection is not restricted to a single imaging technology, but rather denotes a family of task-specific computational pipelines anchored to the available modality.

2. Image preprocessing and anatomical localization

In the ophthalmoscopic pipeline, preprocessing is designed to standardize image appearance and amplify retinal detail before segmentation. Original RGB images are size 1536×1152 px and are cropped to 90% of the visible fundus radius to remove camera-generated labels and edge inhomogeneities. The red channel is discarded due to oversaturation, the blue channel is underweighted by a factor of 0.4, and the green channel is used at full strength, yielding

$I_0(x,y) = 1.0\cdot G(x,y) + 0.4\cdot B(x,y).$

A global histogram equalization is then applied to $I_0$, and contrast stretching may be used so that the 1% and 99% percentiles map to $[0,255]$. Background is estimated by Gaussian blur,

$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$

followed by detail enhancement through subtraction,

$I_1 = I_0 - I_{\mathrm{bg}}.$

Small $\sigma_b$ values, such as 3–5 px, highlight exudates and hemorrhages, while larger values, such as 15–20 px, emphasize the yellow macula. Overexposed regions near image borders are detected by simple brightness thresholds; if more than 30% of the fundus area is lost to cropping, the image is discarded (Cisek et al., 2023).

Optic disc localization is then performed by thresholding the preprocessed image $A=I_1$ at level $T_{od}$ to obtain

$$B_{od}(x,y)= \begin{cases} 1 & \text{if } A(x,y)\ge T_{od}\ 0 & \text{otherwise.} \end{cases}$$

Classical morphology is used with a structuring element $B$, typically a disk of radius $I_0$0:

$I_0$1

After an open-then-close sequence and connected-component labeling, the candidate region whose centroid is brightest in the original $I_0$2 and lies sufficiently far from image borders is selected as the optic disc (Cisek et al., 2023).

Other non-deep pipelines in the literature implement similar anatomical standardization with different primitives. One system extracts the green channel, applies CLAHE, uses Alternate Sequential Filtering with elliptical structuring elements of sizes $I_0$3, $I_0$4, $I_0$5, and $I_0$6 for background estimation, and localizes the optic disc by k-means clustering plus normalized cross-correlation template matching, achieving correct localization in 80/81 images, or 98.77% (Basu et al., 2022). The convergence of these approaches around contrast enhancement, background suppression, and anatomical masking reflects a stable design principle in lesion-first DR analysis.

3. Lesion extraction and diagnostic decision logic

Once the optic disc and macula are excluded, the algorithm detects three lesion classes: microaneurysms and small hemorrhages, blood vessels for hemorrhage refinement, and hard exudates or soft exudates. Dark lesions are first obtained by thresholding $I_0$7 at a low level $I_0$8:

$I_0$9

Connected components are labeled, and for each component the area $[0,255]$0 and circularity

$[0,255]$1

are computed. Components with $[0,255]$2 px$[0,255]$3 and $[0,255]$4 are classified as microaneurysms; larger or elongated components are marked as hemorrhages if they do not overlap the vessel mask (Cisek et al., 2023).

Blood-vessel segmentation is used as a corrective step rather than only as a separate endpoint. A matched-filter bank of oriented Gaussian kernels at $[0,255]$5 is convolved with $[0,255]$6; the maximum response at each pixel gives a vessel probability map $[0,255]$7, thresholded at $[0,255]$8 to obtain the vessel mask $[0,255]$9. Any dark blob overlapping $I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$0 by more than 30% of its area is reclassified as vessel fragment rather than hemorrhage (Cisek et al., 2023). This explicit overlap rule is important because it uses vessel topology to suppress a common false-positive mode in dark-lesion detection.

Bright lesions are obtained by high-intensity thresholding,

$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$1

followed by exclusion of the optic disc, macula, vessels, and hemorrhage regions. Remaining bright connected components are clustered; those with mean distance to the macula center less than $I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$2 and with cluster size greater than 50 px$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$3 are flagged as hard exudates (Cisek et al., 2023).

The final decision logic is rule-based: $I_1 = I_0 - I_{\mathrm{bg}}.$1 with output comprising the diagnosis and overlaid lesion masks (Cisek et al., 2023). This architecture differs from end-to-end image-level classifiers because the diagnostic state is explicitly mediated by lesion counts and lesion classes.

Deep learning systems often preserve this lesion-centric structure even when replacing handcrafted extraction with learned detectors. DRDrV3 uses Mask R-CNN with a ResNet+FPN backbone, reduced anchor sizes with minimum 8 px, and heads for lesion type, bounding box, and mask, followed by a bidirectional LSTM that consumes class, box, and mask-derived features to predict severity classes 0, 1, and 2 (Shenavarmasouleh et al., 2021). DRG-Net goes further by jointly learning multi-lesion segmentation for MA, HE, EX, and SE and five-class DR grading, while also permitting expert correction of lesion masks and grade labels in a weakly supervised interactive loop (Tusfiqur et al., 2022). These systems indicate that lesion-first reasoning remains structurally influential even when the detectors themselves are neural.

4. Application architecture and clinical workflow

The application layer in the target work is a conventional client-server deployment designed for cooperating ophthalmology offices. The client is a Windows-based GUI implemented in C#/.NET WinForms. The user logs in, selects a patient, and uploads fundus images via HTTPS to a central server. The server is a Linux machine running a RESTful service in Python/Flask that invokes a C++/OpenCV DR-analysis executable; results, including lesion masks and DR classification, are returned as JSON and may be rendered as SVG overlays (Cisek et al., 2023).

The workflow is explicitly staged. Right- and left-eye images are selected or drag-dropped; the user presses “Analyze”; progress logs record timestamp, side, and user; and the results page displays the fundus photo with color-coded overlays, numerical counts of each lesion type, and the final DR grade such as No DR or Mild NPDR. The system also provides export to PDF report or DICOM SR to PACS. The listed technologies are C#, .NET 4.7, and OpenXML SDK on the client; Ubuntu 18.04, Python 3.7 with Flask, PostgreSQL, and C++ OpenCV 3.4 on the server; secure communication uses TLS 1.2 and OAuth2 authentication (Cisek et al., 2023).

This implementation pattern situates the diagnostic algorithm within an operational clinical transaction rather than a stand-alone research script. A related browser-based application, DiaRet, exposes retinal image upload, five-grade prediction, class probabilities, and an Integrated Gradients attribution mask through a web interface, emphasizing that deployment can be thin-client as well as desktop-client depending on institutional requirements (Patel et al., 2021).

A plausible implication is that practical DR applications must solve three distinct problems simultaneously: image quality handling, algorithmic lesion analysis, and clinical systems integration. The target application addresses these by combining artifact rejection, morphology-based diagnosis, and transport/export functions suited to ophthalmology offices and PACS-linked workflows.

5. Reported performance and methodological comparisons

The target algorithm was evaluated with five-fold cross-validation against expert gradings. Cohen’s weighted kappa $I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$4 was used for agreement on DR severity levels, and the algorithm achieved $I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$5 on the training set; the same report notes literature values of approximately $I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$6 for inter-expert agreement. Additional statistics reported in extended internal tests were accuracy = 92%, sensitivity = 90%, specificity = 93%, and AUC = 0.95, with the standard definitions

$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$7

and

$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$8

These values place the morphology-based application in the range of clinically meaningful screening performance for the stated dataset (Cisek et al., 2023).

The literature reports a broad spread of results, partly because the tasks differ: vessel segmentation, lesion instance detection, binary DR screening, severity grading, proliferative DR detection, or referral. Representative examples are summarized below.

Study	Method / target	Reported result
(Cisek et al., 2023)	Morphological lesion analysis on 647 hand-held ophthalmoscope images	$I_{\mathrm{bg}}(x,y)=I_0(x,y)*G_{\sigma_b}(x,y),$9, accuracy 92%, sensitivity 90%, specificity 93%, AUC 0.95
(Basu et al., 2022)	Vessel segmentation, optic disc localization, exudate detection, binary DCNN diagnosis	Vessel accuracy 95.93%; optic disc localization 98.77%; DR diagnosis 75.73%
(Shenavarmasouleh et al., 2020)	Mask R-CNN lesion features with neural severity classifier	Test accuracy 92.55%
(Shenavarmasouleh et al., 2021)	Mask R-CNN lesion detection plus bidirectional LSTM severity classification	Best severity accuracy 93.47%; test mAP@0.50 = 0.4370
(Li et al., 2023)	Hierarchical fusion of structural OCT, OCTA, and LSO for PDR vs non-PDR	AUC 0.911, sensitivity 0.86, specificity 0.88
(Moya-Sánchez et al., 17 Aug 2025)	Responsible AI referral system on local dataset of 1,046 patients	Referable DR per patient: F1 98%, accuracy 96%, specificity 96%

These results should not be read as a single leaderboard because the endpoints are not identical. A binary “refer” decision on patient-level macula-centered fields is not directly equivalent to five-class grading, lesion instance segmentation, or multimodal proliferative DR detection. Even so, several structural contrasts are clear. Hybrid lesion-based systems can reach severity accuracy above 92% when lesion localization is used as an intermediate representation (Shenavarmasouleh et al., 2020, Shenavarmasouleh et al., 2021). Multimodal fusion can improve PDR detection relative to the best single modality, with hierarchical fusion increasing AUC from 0.859 for structural OCT to 0.911 when OCTA and LSO are added (Li et al., 2023). Responsible referral systems can prioritize specificity and fairness under deployment constraints, as seen in the comparison against the FDA-approved EyeArt system (Moya-Sánchez et al., 17 Aug 2025).

6. Interpretability, robustness, fairness, and limitations

Interpretability is addressed in the target system through lesion overlays and explicit lesion counts, but related work makes this theme more explicit. DiaRet uses Integrated Gradients to generate attribution masks, where red and yellow regions correspond to pixels most influential for the predicted DR grade, and clinically high-attribution areas correspond to microaneurysms, hemorrhages, hard exudates, and soft exudates (Patel et al., 2021). DRG-Net combines lesion masks, low-level attention maps, and Grad-CAM, and allows an ophthalmologist to correct lesion masks or grading labels so that the system can be updated as a whole, even in a weakly supervised manner (Tusfiqur et al., 2022).

Recent systems also make uncertainty and fairness first-class outputs. DRetNet introduces a multi-stage classifier with Monte Carlo Dropout, defines predictive mean and predictive variance over 30 stochastic forward passes, reports Expected Calibration Error of 0.045 versus 0.12 for a baseline CNN, and provides uncertainty heatmaps in which high-variance regions correspond to artifacts, poor illumination, or subtle lesions (Okuwobi et al., 1 Sep 2025). RAIS-DR evaluates subgroup equity with Disparate Impact and Equal Opportunity Difference; reported DI values range from 0.984 to 1.031 and EOD values are all close to zero, at most $I_1 = I_0 - I_{\mathrm{bg}}.$0, across sex, eye projection, laterality, and age group (Moya-Sánchez et al., 17 Aug 2025).

The main limitations reported across the literature are heterogeneous acquisition conditions, incomplete external validation, and task specificity. DRDrV3 notes that heterogeneity of fundus acquisitions still poses challenges and that Mask R-CNN struggles with the smallest microaneurysms, with IoU @ 0.75 below 0.21 (Shenavarmasouleh et al., 2021). The multimodal PDR study reports moderate dataset size, no external validation beyond a single-center cohort, and unspecified preprocessing, registration, and data augmentation details (Li et al., 2023). RAIS-DR is trained solely on DR and ungradable cases and cannot detect other retinal pathologies such as glaucoma or age-related macular degeneration (Moya-Sánchez et al., 17 Aug 2025).

A recurrent misconception is that automated DR diagnosis is exhausted by a single end-to-end classifier. The surveyed work shows a more heterogeneous field: explicit morphology and rule-based decision logic (Cisek et al., 2023), lesion-instance segmentation plus severity modeling (Shenavarmasouleh et al., 2021), joint segmentation-classification with expert interaction (Tusfiqur et al., 2022), multimodal fusion (Li et al., 2023), and fairness-aware referral pipelines (Moya-Sánchez et al., 17 Aug 2025). This suggests that the central design question is not merely whether to use “AI,” but how to combine image enhancement, anatomical priors, lesion evidence, calibration, and clinical workflow constraints into a diagnostically reliable application.