acceptodds
Withdrawn from ICLR 2027

RetinaPack: Few-Shot Continual Retinal Pre-Screening with Offline Neuromorphic Inference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Retinal pre-screening in rural and resource-constrained settings requires portable, offline and privacy-preserving inference that can adapt to local acquisition shifts without large-scale retraining. A key difficulty is that current-stage labels can be obtained immediately, whereas prognostic outcomes mature months later, so local adaptation must improve staging without destabilizing longitudinal risk estimates. We introduce RetinaPack, a participant-disjoint longitudinal benchmark of 13,040 participants, and RetinaLPA-2S, a few-shot continual adaptation method for four-stage age-related macular degeneration (AMD) pre-screening. RetinaPack combines adjudicated change and stable pairs with crossed camera/operator repeats to separate disease-related variation from acquisition nuisance and define a pathology-referenced protected subspace. RetinaLPA-2S updates only its orthogonal complement through a 4,096-parameter two-sided low-rank adapter, while a frozen competing-risk branch preserves fixed-input prognosis during stage-only adaptation. A quantized encoder runs on the Akida AKD1000 and signed adaptation remains on the local host, enabling network-independent inference, local updating and feature replay without retinal-image storage. With 10 labels per stage, RetinaLPA-2S reaches 82.20 domain-balanced candidate F1 with 1.07-point forgetting, 1.68 points above BiLoRA-style FA. On 920 staging queries, it achieves 83.78 macro-F1 and 95.33% referral sensitivity. The complete image-to-report path requires 2.150 J, 650.4 ms P95 latency and 224 MiB peak memory, reducing processing energy by 34.35% relative to CPU INT8.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
retinal pre-screening, age-related macular degeneration, continual learning, few-shot adaptation, parameter-efficient adaptation, neuromorphic inference, offline inference, longitudinal risk prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NwMaOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RetinaLPA-2S, a few-shot continual adaptation method for retinal disease screening, along with Retinapack, a retinal imaging benchmark for longitudinal analysis. The core idea is to identify and protect a subspace (low-dimensional) associated with pathology changes while allowing the model to adapt to new conditions. The paper also explores deployment on resource-constrained hardware.

Critical strengths
  • The paper works on a very important problem, and its healthcare relevance cannot be overstated, especially in rural and resource-constrained settings where access to ophthalmologists is limited. As the paper points out, development of reliable AI systems that can adapt (or continually learn) and can operate offline could have meaningful impact on early disease screening and clinical care. Here early disease screening could be for the eye disaease and this could also be applicable for other systematic diseases (i.e., oculomics).

  • The paper introduces RetinalPack which is a valuable contribution that could support future AI research in retinal imaging. This benchmark includes more than 13K participants, longitudinal disease-change and stable pairs, repeated acquisitions across cameras (also operators), and separate cohorts for adaptation and evaluation. The longitudinal aspect is particularly interesting as I'm not aware of any such benchmark. This will allow the studies of disease progression beyond conventional cross-sectional classification/detection.

Critical weaknesses
  1. The paper is strongly motivated by an important clinical application, but I'm not fully convinced about the strength of its methodological contribution for ICLR. The proposed approach combines generalized eigenvalue-based subspace identification, low-rank adaptation, and existing continual-leanring techniques, and while overall framework is interesting, it is not entirely clear what fundamentally new AI methodology is being contributed here.

  2. The central methodological contribution of subspace identification techniques is used to identify a low-dimensional subspace associated with longitudinal pathology changes. But its unclear if the identified subspace genuinely captures pathology-related changes. The authors seem to construct it using clinically labeled change and stable pairs, but I'm not sure if the experiments directly demonstrate that these directions correspond to meaningful pathological changes rather than other correlated variations.

  3. Related to point 2, another concern is whether the identified subspace genuinely captures semantic changes in data and if the proposed method can be generalized to other datasets and settings.

  4. The question of generalization is also relevant for the presented continual learning and parameter efficient approach. While the paper made comparison with existing works, these comparisons are primarily conducted on the authors' own retinal benchmark.

  5. Finally, the paper emphasize offline deployment but the evaluation is largely based on controlled hardware experiments rather than actual clinical deployment. I can understand that the acutual clinical deployment is perhaps out of scope of this study but since deployment is presented as contribution, I think this warrant the understanding of how well the benefits would hold under different hardware configurations and real-world operating conditions.

Questions for authors
  1. Please answer the points raised in the weakness section.

  2. I am also curious about the clinical acceptance criteria. I see that the paper establish specific criteria before an adapted model can be released. But there is also report about how well the model correctly detects C3 cases but still fails under predefined confidence-bound requirement. So my question is how often can the proposed continual learnign/adaptation system actually deploy its updated models if we see failure in terms of clinical acceptance criteria?

Reviewer_srf9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents RetinaPack, a longitudinal retinal benchmark, and RetinaLPA-2S, a few-shot continual adaptation system for retinal pre-screening. The method estimates a pathology-referenced subspace from adjudicated longitudinal pairs and repeated acquisitions, and restricts local staging updates to its orthogonal complement. The system further includes a frozen competing-risk branch, feature replay, release gates, rollback, and hybrid neuromorphic execution.

The paper addresses an important problem and demonstrates unusually careful experimental governance. However, I do not currently find the central machine-learning contribution sufficiently established for acceptance. In particular, the algorithmic novelty is modest, several key baselines are feature-space reductions rather than native implementations, and the headline candidate-only results do not represent the gate-aware retained model that the deployment claims depend on.

Critical strengths
  1. The clinical motivation is meaningful and clearly articulated.

  2. The participant-level data separation, multiple domain orders, paired inference, release validation, and rollback design are unusually careful.

  3. The authors distinguish candidate checkpoints from retained states and disclose a clinically relevant gate failure rather than hiding it.

  4. The hardware evaluation reports latency, energy, and memory under a shared processing boundary.

Critical weaknesses
  1. The central algorithmic novelty is limited. The two-sided projection provides clear algebraic invariants, but its construction is relatively direct. Moreover, the appendix shows that under suitable rank conditions a direct readout parameterization can realize the same logit corrections. Thus, the factorized adapter does not establish a fundamentally different function class. The primary novelty appears to lie in the clinically constructed subspace, but its broader methodological significance remains unclear.

  2. The baseline evidence does not support the apparent breadth of the claims. Several important comparisons, including BiLoRA-style FA and GPM-FA, are matched feature-space reductions rather than native implementations. Although this controls the trainable parameter budget, it does not establish superiority over the original methods under their intended parameterizations and optimization procedures.

  3. The headline continual-learning result is not aligned with the proposed deployment procedure. The main table emphasizes candidate-only trajectories. However, the proposed system applies release gates and rollback, and at least one clinical-domain candidate fails the prespecified sensitivity-bound gate. A rejection changes the retained parent, teacher, and subsequent optimization trajectory. Therefore, candidate-only performance is not an adequate proxy for the actual deployed continual learner. The complete gate-aware retained-state trajectory in the primary clinical configuration should be the main experimental result.

  4. Several preservation claims are stronger than the demonstrated guarantees. Preserving coordinates in a learned subspace does not demonstrate preservation of all pathology-relevant semantics. Freezing a bypassed risk branch provides fixed-input output invariance, but not prognostic validity under domain shift. Similarly, feature-only replay reduces raw-image retention but does not establish privacy against feature reconstruction or inference attacks.

  5. The breadth of the system obscures the core ICLR contribution. The paper combines benchmark construction, continual adaptation, risk prediction, clinical workflow analysis, release governance, privacy arguments, and hardware deployment. Each component is potentially valuable, but the combined presentation does not compensate for the lack of a clearly isolated and convincingly compared ML advance. Static staging is not significantly better than the strongest foundation-model probe, the compact risk model is not the best-performing risk model, and the hardware path is a hybrid accelerator–host deployment.

Questions for authors
  1. Can the authors report the full gate-aware retained-state trajectory for the primary clinical configuration, including all re-optimized post-rejection updates?

  2. How does the method compare with native implementations of the strongest continual-learning and PEFT baselines under matched supervision and computational budgets?

  3. What empirical result demonstrates that the benefit comes specifically from the clinically estimated subspace rather than replay, anchoring, optimization, or the chosen parameterization?

Reviewer_htHXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces two contributions: RetinaPack, a longitudinal, non-overlapping, participant-level retinal pre-screening benchmark (comprising 13,040 individuals, including adjudicated pairs of "changed" vs. "stable" cases, and a sub-study of 240 individuals with repeated acquisitions using two different cameras and operators); and RetinaLPA-2S, a method for few-shot continual adaptation. RetinaLPA-2S employs bilateral low-rank adapters (with 4,096 parameters) that operate outside a "protected subspace"—a subspace derived from the second-order moments of "pathological change" versus "stability/acquisition noise." Additionally, the prognostic branch utilizes a frozen competing-risk head with a bypassed adapter, ensuring that risk outputs remain invariant during localized updates focused solely on disease staging.

Critical strengths
  1. The question of "which coordinates should be protected" is well-posed, and the chain of evidence is more comprehensive than in comparable studies. While existing PEFT/CL methods typically define protection constraints based on weights, logits, gradients, or general parameter geometry, the authors take a different approach: they anchor the protected subspace to the direction of longitudinal clinical pathology, estimating this using three types of matched pairs—adjudicated change pairs, stable pairs, and repeat-acquisition pairs.

  2. The evaluation protocol approaches the rigor of clinical research, with the authors demonstrating transparency by reporting findings that are unfavorable to their own work. The methodology includes non-overlapping splits at the participant, eye, and visit levels; independent dual-scoring with specialist adjudication; strict isolation between support, release, and query sets; pre-specification of release thresholds prior to freezing candidate models; a rollback mechanism for failures with a full audit trail; and versioned hashing for both the encoder and the risk-prediction head. Furthermore, the inclusion of a real-world clinical workflow analysis (17.5% re-acquisition rate, 3.5% manual escalation, and 96.5% automated output coverage), combined with participant-level paired inference, provides a far more compelling validation than the standard practice of simply reporting a macro-F1 score.

Critical weaknesses
  1. The core advantages claimed only hold true when using the authors' custom-developed, ultra-small encoder (24,512 parameters); they do not translate to the strong baselines cited in the paper. The entire set of comparisons for continual learning (the "FA" series in Table 3) is based exclusively on this encoder, which was trained from scratch: 82.20 vs. 80.52 for BiLoRA—a margin of 1.68 points. Given that the standard deviation (SD) across training runs ranges from 1.31 to 1.72, this difference is statistically marginal; the authors themselves acknowledge that these figures are derived from specific, fixed training runs and do not account for stochasticity. Furthermore, for the staged tasks, the authors report that models like RETFound, VisionFM, and RetiZero achieve F1 scores of 85.0–86.1 when fine-tuned via PEFT or full fine-tuning—significantly outperforming the 83.78 F1 score reported for the proposed method. In other words, in terms of "practically usable accuracy," the method is currently outperformed by foundational models subjected to simple fine-tuning. Additionally, Appendix D10 provides results for DER++/GPM/BiLoRA/CODA-Prompt using a standard ViT-B/16 backbone, yet it omits results for RetinaLPA-2S on any ViT backbone—a glaring omission in the paper. If the gains from "bilateral constraints" are diluted when applied to a stronger backbone—due to the inherent adaptation capabilities of the base model—then the paper’s central argument regarding continual learning effectively collapses.

  2. The comparisons provided in Appendix Q largely negate the methodological contribution of "bilateral factorization." Moreover, the performance advantage of the "protected subspace" over the "general subspace" is marginal and lacks statistical significance. In the analysis of the feature space trajectory, a direct control approach—utilizing 1,024 matrix entries and an induced metric penalty—achieves an F1 score of 90.69 with a forgetting score of 1.17, performing identically to the bilateral factorization approach (which uses 4,096 entries). The authors subsequently concede that the true driver of performance is likely a combination of "clinical subspace selection" and an efficient trade-off between parameterization and optimization. In other words, the unique value of the bilateral structure itself is largely negated; the paper’s actual contribution effectively narrows down to "selecting a protected subspace via clinical pairing." Moreover, the magnitude of this advantage is limited: 82.20 (clinical pairing) versus 79.94 (random selection) and 80.76 (gradient weighting)—a difference of 1.3–2.3 points. This is on the same order of magnitude as the trajectory standard deviation (SD), and the appendix explicitly states that "trajectory SD alone is insufficient to establish significance at the participant level." Additionally, setting the protected rank to 16 (yielding 82.37) performs slightly better than setting it to 8.

Questions for authors
  1. When and in what format will RetinaPack be made publicly available? If the data remains inaccessible during the review process, reviewers cannot verify the quality of the annotations for "change pairs," "stable pairs," and "repeated acquisition pairs," nor can they assess the design of the sub-studies. Consequently, they would have to accept the entire chain of evidence based solely on the figures reported by the authors—a situation that would significantly lower my recommendation.
  2. Could you provide the continual learning results for RetinaLPA-2S using a standard, real-world backbone (such as RETFound or ViT-B/16), compared against methods like BiLoRA or CODA-Prompt under identical conditions—specifically with matching support labels, replay budgets, and data access protocols? I would like to determine whether the 1.68-point performance gain stems from an effective selection of subspaces or from the fact that the smaller encoder is simply easier to optimize.
  3. In Appendix Q, the "direct readout control" performs identically to the "bilateral factorization" approach. Does this imply that the method should be reframed as "a clinically protected subspace combined with an arbitrary, parameter-efficient parameterization of the complementary subspace"? If so, can the bilateral factorization still be considered a primary algorithmic contribution?
  4. Given a realistic thresholding scenario—where 40% of candidates are rejected and the entire D3 domain is excluded—how close are the reported figures (82.20 / 83.78) to the performance that would be achieved in an actual clinical deployment? Furthermore, could you provide the Macro-F1 scores across domains D1–D3 based on the "gate-aware" (retained) trajectories, rather than those derived from the reference feature space?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.