i am doing a research on diabetic retinopathy detection. the current research gap we have is to train an evidential model that detects DR; this model will be trained only on DR fundus images. we will feed this model fundus images of other eye diseases like glaucoma, hypertensive retinopathy, cataracts. what our model will do is trigger an OOD rejection if these fundus images are given instead of giving a forced DR grading (0 to 5).  can you search up past studies and let me know if this gap still exists and whether i can actually start working on it or not this research gap is just a generalized description; make improvements accordingly such that the gap will actually be impactful and will have less risks once i start working on it

Short answer

Yes—this is worth pursuing, but not as the claim that “evidential learning can make a DR grader reject other eye diseases.” That broad gap is already substantially occupied. Prior work has used Dirichlet Prior Networks to identify distribution shifts and OOD inputs in DR screening [citation_272do8], reported better near-OOD detection of samples containing other eye diseases using deep-kernel learning for DR grading [citation_272dnq], and developed retinal open-set systems that assign uncertainty to unseen retinal-disease categories, low-quality fundus images, and non-fundus images [citation_272e18]. A separate anomaly-detection study evaluated glaucoma, cataract, hypertensive retinopathy and other retinal abnormalities [citation_272dpu]. Thus, the problem is real, but the headline method/problem combination is not new by itself.

A narrower, potentially publishable gap remains: a rigorous, clinically framed head-to-head evaluation of whether an evidential five-grade DR model, trained only on DR-grading data, can both preserve in-distribution grading and safely abstain on previously unseen, same-modality ocular diseases—with disease-held-out, external-site testing and calibrated referral thresholds. In the abstracts found in this search, I did not find a study that clearly establishes that exact combination and evaluation. Treat this as a candidate gap, not a confirmed novelty claim: the search was abstract-based and one provider had an incomplete search lane.

What the closest studies already establish

• DR-specific OOD detection is not new. Nandy and colleagues describe a Dirichlet Prior Network framework with an OOD detector plus a DR classifier, identifying both non-retinal inputs and distribution-shifted retinal images for human review [citation_272do8].
• Other eye diseases are already a known “near-OOD” challenge for DR models. Siebert and colleagues compared uncertainty methods for DR grading and report that deep-kernel learning improved detection of near-OOD samples containing other eye diseases [citation_272dnq]. This is especially close to your intended use case.
• Open-set rejection for retinal disease more broadly already exists. The 2023 UIOS study trained on fundus images from nine retinal conditions and reports higher uncertainty for unseen retinal-disease categories, low-quality images and non-fundus inputs [citation_272e18]. Its abstract reports F1 of 91.91% on a test set of unseen categories, compared with 64.74% for a standard AI model; those figures are the paper’s reported results, not directly comparable to DR-only grading [citation_272e18].
• Anomaly detection has been tested against your named diseases. Han and colleagues evaluated a fundus anomaly detector across DR, glaucoma, cataract, hypertensive retinopathy, AMD, myopia and other abnormalities, with external test sets [citation_272dpu]. It is not the same as a DR-only evidential grader with a reject option, but it makes a simple “other diseases are unstudied OODs” claim untenable.
• Uncertainty/rejection for DR, without your precise OOD focus, is also being explored. A 2025 preprint evaluates rejection based on uncertainty and reports a coverage–reliability trade-off, using a variational Bayesian approach [citation_272dp3].

Make the gap sharper and more impactful

A stronger framing is:

Can evidential uncertainty support safe open-set referral in ordinal DR grading when the model has never been trained on other ocular diseases? Compare an evidential DR grader against strong non-evidential and uncertainty baselines, and test whether its rejection decisions generalize to held-out ocular diseases and external imaging sites without sacrificing DR grading performance.

This avoids claiming novelty for “evidential learning + OOD” alone. Your contribution becomes a controlled comparison and clinically relevant validation: (a) the distinction between unseen disease and ordinary image-quality/domain shift; (b) class-wise performance on specific diseases; and (c) whether rejected cases are actually sent to human review at a useful rate. The evidential head should be treated as a hypothesis to test—not assumed to be better. Prior DR work already studies other-eye-disease OOD detection with deep-kernel learning [citation_272dnq], and open-set retinal work already uses uncertainty-based rejection [citation_272e18].

Suggested research question and feasible design

Research question: Among fundus photographs with a reference DR grade, does an evidential DR grader trained only on DR-grading images detect unseen ocular diseases more effectively than standard softmax and established uncertainty/OOD baselines, while maintaining acceptable grading performance on DR images?

Proposed design (not a claim from the literature):

1. Define the task precisely. Train a closed-set grader on all intended DR grades, including the no-DR grade if the goal is to return the lowest grade for unaffected images. If “only DR images” literally means only images with positive DR, the model cannot learn the no-DR grade. Make clear what you mean by “0–5”: it denotes six labels. Some retrieved studies instead describe a five-category grading scale [citation_21yrdk]. Name the scale and reference standard explicitly.
2. Keep training and OOD test genuinely separate. Train only on DR-grading images. Reserve glaucoma, cataract, hypertensive retinopathy and other non-target diseases for evaluation; use separate development/threshold-calibration and final test sets. Add disease-held-out tests (e.g., tune/validate on some non-DR conditions, test on a condition excluded from that process) to assess generalization beyond memorized known outliers. Report each disease separately, not only a pooled OOD score.
3. Separate OOD from related but different failure modes. Include (i) other ocular disease, (ii) camera/site/domain shift, (iii) ungradable or poor-quality retinal images, and (iv) non-fundus images. These groups are not interchangeable: previous work treats disease shifts, image quality and non-fundus inputs as distinct review-triggering cases [citation_272e18]. Include mixed DR-plus-comorbidity images where labels permit; these are not simply “non-DR” and could be wrongly rejected by a binary disease rule.
4. Compare fairly. Use the same image splits and backbone for an evidential model and conventional softmax maximum-probability/entropy, energy-based OOD, and at least one strong uncertainty baseline such as deep-kernel learning or an ensemble. The DKL comparison is important because it has already been applied to other-eye-disease near-OOD in DR grading [citation_272dnq]. If implementation and label access permit, consider the 2023 retinal open-set method as a contextual comparator [citation_272e18]. If the tasks cannot be aligned, explain the mismatch rather than implying a direct benchmark.
5. Evaluate both utility and safety. Report DR grade performance (e.g., macro-F1 and quadratic-weighted kappa), OOD AUROC and AUPR, FPR at a prespecified high OOD sensitivity, per-disease sensitivity, calibration, and selective risk versus coverage (accuracy/error among accepted cases at different referral rates). Select the rejection threshold on a validation set, then freeze it before external testing. A single AUROC or accuracy figure is not enough to establish safe rejection.
6. Protect against leakage and shortcuts. Split by patient and, where possible, by acquisition site/device; check that disease labels are not confounded with dataset source, image quality, or preprocessing. Use an external test set and report uncertainty intervals. Verify disease labels and DR grades against reliable annotation, and distinguish image-level from patient-level outcomes.

Practical go/no-go: Start if you can obtain (or assemble) patient-level DR grading data plus independently labelled non-DR disease images, with enough samples from more than one source to evaluate site and disease shift. Before full development, run a small pilot: first reproduce a DR grader, then test simple confidence-based rejection on a disease-held-out set. Proceed with evidential modeling only if it offers an improvement over these baselines at a fixed referral/coverage level. If you have only one small, single-source dataset, narrow the project to a reproducible benchmark study and avoid claims of clinical readiness.

Bottom line

Proceed with a revised, evidence-comparison-focused project; do not present the broad concept as an untouched research gap. The strongest defensible contribution is testing whether evidential uncertainty adds measurable value over prior DR OOD/open-set approaches for unseen ocular diseases, under strict disease-held-out and external-site evaluation, while retaining usable DR grading. The question of whether the exact formulation has already appeared remains not fully settled by this search: I searched selected academic databases only, in abstract mode; no full texts were retrieved, and OpenAlex had an incomplete retrieval lane in the first search. I therefore cannot certify priority or comprehensively rule out a closely matching paper. A full-text check and a targeted exact-title/DOI search are appropriate before locking the novelty claim.